Sleeping Beauty and De Nunc Updating
Transcript of Sleeping Beauty and De Nunc Updating
University of Massachusetts Amherst University of Massachusetts Amherst
ScholarWorks@UMass Amherst ScholarWorks@UMass Amherst
Open Access Dissertations
5-2010
Sleeping Beauty and De Nunc Updating Sleeping Beauty and De Nunc Updating
Namjoong Kim University of Massachusetts Amherst
Follow this and additional works at: https://scholarworks.umass.edu/open_access_dissertations
Part of the Philosophy Commons
Recommended Citation Recommended Citation Kim, Namjoong, "Sleeping Beauty and De Nunc Updating" (2010). Open Access Dissertations. 242. https://scholarworks.umass.edu/open_access_dissertations/242
This Open Access Dissertation is brought to you for free and open access by ScholarWorks@UMass Amherst. It has been accepted for inclusion in Open Access Dissertations by an authorized administrator of ScholarWorks@UMass Amherst. For more information, please contact [email protected].
SLEEPING BEAUTY AND DE NUNC UPDATING
A Dissertation Presented
by
NAMJOONG KIM
Submitted to the Graduate School of the
University of Massachusetts Amherst in partial fulfillment
of the requirements for the degree of
DOCTOR OF PHILOSOPHY
May 2010
Philosophy
SLEEPING BEAUTY AND DE NUNC UPDATING
A Dissertation Presented
by
NAMJOONG KIM
Approved as to style and content by:
Phillip Bricker, Chair
Hillary Kornblith, Member
Christopher Meacham, Member
Lyn Frazier, Member
Phillip Bricker, Department Head
Philosophy
iv
ACKNOWLEDGMENTS
First and foremost, I would like to thank my advisor, Phillip Bricker. Ever since I
came up with a small idea about how to update one’s degrees of self-locating beliefs, he
patiently helped me to develop it into a sophisticated theory as presented in this
dissertation. During that process, he continued to offer lots of encouragement, advice, and
sometimes criticisms. Truly, this dissertation would not have been born without his help.
I am greatly indebted to the other members of the committee as well. Through
his teaching and writing, Hillary Kornblith taught me how to philosophize basic
epistemological problems in a clear, effective way. Thanks to him, I better understood the
connection between many issues of mainstream and formal epistemologies. In particular,
he helped me identify the conditions that ought to be satisfied by any cogent rule for
updating.
I cannot overemphasize the important role Chris Meacham played in the early
development of my thoughts. Only after I read his “Sleeping Beauty and the Dynamics of
De Se Beliefs,” did I realize that a new rule for updating the degrees of self-locating
beliefs was essential for the right solution of the Sleeping Beauty problem. After he came
to UMass, we enjoyed a good number of discussions about this dissertation and many
other issues in the philosophy of probability.
It has been a privilege to have Lyn Frazier on my committee. She gave me
invaluable opportunities to think about philosophical probabilism from an outsider’s
point of view. Although she is a linguist and I am a philosopher, our differences worked
as a catalyst of new ideas rather than a barrier between disciplines.
v
The second chapter of this dissertation was published in Synthese. So its
copyright belongs to Springer-Verlag New York, LLC. I am grateful to the editor of
Synthese, Wiebe van der Hoek, for allowing me to include it as a chapter in my
dissertation.
Finally, I would like to thank the other faculty and graduate students in the
philosophy department at UMass. I enjoyed every bit of the seminars and discussions in
this department. Especially, I want to express my gratitude towards Kirk Michaelian, who
became my closest friend.
vi
ABSTRACT
SLEEPING BEAUTY AND DE NUNC UPDATING
MAY 2010
NAMJOONG KIM, B.A., SOGANG UNIVERSITY
M.A., SOGANG UNIVERSITY
Ph.D., UNIVERSITY OF MASSACHUSETTS AMHERST
Directed by: Professor Phillip Bricker
About a decade ago, Adam Elga introduced philosophers to an intriguing puzzle.
In it, Sleeping Beauty, a perfectly rational agent, undergoes an experiment in which she
becomes ignorant of what time it is. This situation is puzzling for two reasons: First,
because there are two equally plausible views about how she will change her degree of
belief given her situation and, second, because the traditional rules for updating degrees
of belief don’t seem to apply to this case.
In this dissertation, my goals are to settle the debate concerning this puzzle and
to offer a new rule for updating some types of degrees of belief. Regarding the puzzle, I
will defend a view called “the Lesser view,” a view largely favorable to the Thirders’
position in the traditional debate on the puzzle. Regarding the general rule for updating, I
will present and defend a rule called “Shifted Jeffrey Conditionalization.”
My discussions of the above view and rule will complement each other: On the
one hand, I defend the Lesser view by making use of Shifted Jeffrey Conditionalization.
On the other hand, I test Shifted Jeffrey Conditionalization by applying it to various
credal transitions in the Sleeping Beauty problem and revise that rule in accordance with
vii
the results of the test application. In the end, I will present and defend an updating rule
called “General Shifted Jeffrey Conditionalization,” which I suspect is the general rule
for updating one’s degrees of belief in so-called tensed propositions.
viii
CONTENTS
Page
ACKNOWLEDGMENTS……………………………………………………….……….iv
ABSTRACT……………………………………………………………………………...vi
LIST OF TABLES………………………………………………………………………..xi
LIST OF FIGURES……………………………………………………………………...xii
CHAPTER
1. INTRODUCTION .................................................................................................. 1
A. Problem ....................................................................................................... 1
B. Goals ........................................................................................................... 3
C. Strategy ....................................................................................................... 5
D. Contents .................................................................................................... 13
2. UPDATING WITH A SINGLE OBSERVATION .............................................. 20
A. Introduction ............................................................................................... 20
B. Background ............................................................................................... 24
C. A Problem of the De Se Versions of SC and JC ....................................... 27
D. Shifted Jeffrey Conditionalization ............................................................ 30
E. Shifted Rigidity as a Conditional Expert Principle ................................... 37
F. Sleeping Beauty and Shifted Jeffrey Conditionalization .......................... 44
G. Conclusion ................................................................................................ 49
3. UPDATING WITH A SEQUENCE OF OBSERVATIONS ............................... 52
A. Introduction ............................................................................................... 52
B. Review of SC and SJC .............................................................................. 54
C. Strategy ..................................................................................................... 58
D. Updating with a Sequence of Observations .............................................. 60
E. A Defense of SSJC.................................................................................... 75
F. The SB Problem and the Inconsistency of SSJC ...................................... 81
G. A Diagnosis and a Potential Solution ....................................................... 82
H. Conclusion ................................................................................................ 91
4. UPDATING WITH DE PRIORI INFORMATION ............................................. 93
A. Introduction ............................................................................................... 93
B. Strategy ..................................................................................................... 94
C. Updating with De Priori Information ....................................................... 98
E. Temporal Conditional Multiple Expert Principle ................................... 111
F. A Defense of GSJC ................................................................................. 121
G. Too Far Past Does Not Matter ................................................................ 123
ix
H. Application to the SB Problem ............................................................... 127
I. The Relation between GSJC and Other Rules ........................................ 134
J. Conclusion .............................................................................................. 137
5. SATISFACTION OF DESIDERATA ................................................................ 139
A. Introduction ............................................................................................. 139
B. Background ............................................................................................. 140
C. Strategy ................................................................................................... 144
E. GSJC+ and GSR
+ ..................................................................................... 147
F. The Transitivity of GSR+ ........................................................................ 150
G. Synchronic and Diachronic Coherence ................................................... 157
H. Observational Exhaustiveness ................................................................ 162
I. Filling the Gap ........................................................................................ 165
J. Conclusion .............................................................................................. 168
6. CONCLUSION ................................................................................................... 169
A. Summary ................................................................................................. 169
B. Remaining Issue 1: Generalization for De Se Updating ......................... 171
C. Remaining Issue 2: Credence Distribution over the Partition ................ 176
D. Remaining Issue 3: The Possibility of a Rival Rule ............................... 181
E. Conclusion .............................................................................................. 187
APPENDICES
A. EQUIVALENCE BETWEEN GSJC- AND GSJC. ................................................... 189
B. EQUIVALENCE BETWEEN GSR+ AND GSJC
+ .................................................... 194
C. TRANSLATION BETWEEN TEMPORAL CONTEXTS ....................................... 198
D. EQUIVALENCE BETWEEN GSJC+ AND GSJC
0 .................................................. 206
BIBLIOGRAPHY ........................................................................................................... 213
xi
LIST OF FIGURES
Figure Page
1: Update in Accordance with SJC…………………………………...………………….15
2: Update in Accordance with SSJC……………………………………………………..16
3: Update in Accordance with GSJC………………………………………………...…..18
4: Transitivity of GSJC………………………...………………………………………...19
5: Rain and Precipitation…………………………………………………………………42
6: Evidential Uncertainty in Sequential Updating…………...…………………………..71
7: Temporal Uncertainty in Sequential Updating…………………………...…………...73
8: Flying Birds, Running Animals, and Earthquake……………………………………..79
9: Deference by SBR and Temporal Ignorance………...…………………………...…...89
10: Reindexicalization and Deindexicalization 1………………………………………107
11: Reindexicalization and Deindexicalization 2…………………...………………….109
12: Judgmental Dependence………………………………………………...………….111
13: Direct and Indirect Data Providers…………………………………...…………….113
14: Referents of Indexicals in Different Temporal Contexts…………………………...199
15: Indexicals, Observations, and Intervals 1………………………………...………...200
16: Indexicals, Observations, and Intervals 2………………………...………………...201
17: Indexicals, Observations, and Intervals 3………………………...………………...202
1111
CHAPTER I
INTRODUCTION
A. Problem
In his “Self-locating Belief and Sleeping Beauty,” Elga (2000) presents an intriguing
puzzle: On Sunday night, Sleeping Beauty (hereafter: SB) knows that she will go through
the following experiment. On that night, she is put to sleep by a group of evil
experimenters. Then, they toss a fair coin. Case 1: (H) The coin lands heads. In this case,
the experimenters wake her up only once, on Monday. Case 2: (T) The coin lands tails. In
this case, they wake her up twice, the first time on Monday and the second time on
Tuesday. Between the two awakenings, they inject SB with a drug that erases her
memory of the first awakening. In either case, one minute after she wakes up on Monday,
she is told that it is Monday, and, when the experiment ends on Wednesday, she is
awakened with her memory of the last awakening intact. The puzzle ends with two
questions: When SB wakes up on Monday, what is her degree of belief in H? When she is
told that it is Monday, what is her degree of belief in H?1
There have been two dominant answers to the first question in the literature.
Halfers argue that the answer is 1/2 (Lewis 2001; Bradley 2003; Jenkins 2005): On
Monday, she wakes up with Sunday as her last memory. Since she fully expected to wake
up in that way, SB receives no new evidence relevant to H at that moment. Intuitively, a
1 This version of the SB problem is closer to Lewis’s (2001) version because the step of telling SB the day
is omitted in Elga’s original version (2000). In Chapter II, I will discuss Elga’s (simpler) version, to focus
on the first question, and in Chapter III and later, I will use Lewis’s (more complex) version, to answer the
second question.
2222
rational agent changes her degree of belief (hereafter: credence) in a proposition X only
when she receives new evidence relevant to X. SB’s credence in H was 1/2 on Sunday
night. Therefore, her credence in H on Monday is also 1/2. (Lewis 2001, p. 174.)
Thirders, on the other hand, contend that the answer is 1/3 (Elga 2000; Dorr 2002;
Weintraub 2004). According to their view, when SB wakes up on Monday, she knows
that she is waking up on either Monday or Tuesday but does not know which day it is. On
the one hand, she assigns 1/2 to H conditional on the possibility that she is waking up on
Monday. For remember that on Sunday night, she assigned 1/2 to H conditional on her
waking up on Monday. On the other hand, she assigns 0 to H conditional on the
possibility that she is waking up on Tuesday. This is because SB knows that if she is
waking up on Tuesday, the coin has already landed on tails. It follows that her actual
credence on Monday in H will be the weighted average between 1/2 and 0, where the
weights come from her credences that she is waking up on Monday and that she is
waking up on Tuesday. Since she cannot be sure that she is waking up on Monday, her
rational credence in H is less than 1/2. If we take symmetry into consideration, it will be
1/3. (However, many philosophers doubt that symmetry can restrict an agent’s credence
in this way. For the purposes of my argument, the exact value of SB’s credence on
Monday in H is not important. The important element of the Thirders’ view is that SB’s
credence in H is less than 1/2 when she wakes up on Monday.) (Elga 2000, pp. 144-145.)
On the second question, most Halfers and Thirders agree that when she learns
that it is Monday, SB's credence in H increases, but they disagree about the precise value
of the resulting credence in H. Many Halfers believe that it is 2/3, and virtually all
Thirders believe that it is 1/2.
3333
To compare their views, look at Table 1:
Table 1: The Change of SB’s Credence in H
Sunday night Monday morning told “Monday”
Halfers 1/2 1/2 2/3
Thirders 1/2 1/3 1/2
Ever since the publication of Elga’s paper, philosophers have debated between these two
options. So which side is right?
B. Goals
In this dissertation, I pursue two goals: First, I shall offer a solution to the SB problem;
basically, I shall defend the Thirders’ view minus any use of symmetry. Second, I shall
develop a general rule for updating de nunc credences; in other words, I shall discuss a
method for updating degrees of belief in tensed propositions.2
As the SB problem has attracted so many philosophers’ attention, my pursuit of
the first goal hardly requires any explanation. But why do I pursue the second? I do so
because it offers the straightest solution to the SB problem.
To appreciate this point, think about these facts: We know SB’s credence
distribution on Sunday night fairly well. For instance, we know that her credence in H on
that night is 1/2, we know that she knows then that it is Sunday, etc. Thus, finding her
2 By “de nunc credence,” I mean one’s degree of belief in a tensed proposition. By “tensed proposition,” I
mean a proposition-like entity that may have different truth-values relative to time, but not relative to the
individual. For instance, “it is raining now in Boston” may have different truth-values on Monday and
Tuesday, but if it is true/false on Monday for someone, then it is true/false on Monday for everyone.
4444
credence in H on Monday must be simply a matter of applying an appropriate updating
rule to SB’s credal transition from Sunday night to Monday morning. Therefore, if we
have a correct rule for updating applicable to this transition, it will be easy, in principle,
to calculate her credence in H on Monday.
Nevertheless, this approach has not been taken by many philosophers. Why? The
traditional updating rule, called “Strict Conditionalization,” states that an agent’s
credence at t in a proposition X is her conditional credence at t′<t in X given E, where E is
the totality of her observations during (t′,t]. Unfortunately, this rule does not apply to
SB's credal transition from Sunday night to Monday morning. SB learns (W) “SB wakes
up today with the memory of Sunday as the last memory remembered” on Monday.
However, on Sunday SB must have known (W′) “SB woke up today with the memory of
Saturday as the last memory remembered.” Since W is logically incompatible with W′,
her conditional credence on Sunday night in H given W has no defined value. Therefore,
Strict Conditionalization fails to provide a defined value for the former credence.
Hence, the most effective way to solve the SB problem is to apply an appropriate
updating rule to SB’s transition from Sunday night to Monday morning, but the Strict
Conditionalization rule is inappropriate here. While many other philosophers have tried
to solve the problem by appealing to other considerations for this difficulty (e.g. Lewis
appeals to his Principal Principle to settle the debate (Lewis 2001), Kierland and Monton
resort to the principle of minimizing inaccuracy of one's credence (Kierland and Monton
2005), etc.), I believe that finding the correct rule for de nunc updating is the best way to
settle the debate.
5555
In summary, I pursue two goals: Solving the SB problem and finding the general
rule for de nunc updating. I have suggested that achieving the second goal will be the
quickest way to achieve the first, but this does not mean that the second goal only has a
derivative value. Indeed, it is the opposite: The SB problem is so interesting precisely
because it reveals the fact that a new rule for updating is necessary to deal with de nunc
credences properly.
C. Strategy
As I said above, I intend to develop a new rule for updating de nunc credences. To this
end, I have developed the following criteria for an acceptable rule:
Solution I want my updating rule to provide intuitive answers to the two
questions of the SB problem.
Versatility I want my updating rule to apply to as many types of updating
situations as conceivable.
Coherence I want my updating rule to provide coherent results.3
How can I develop a de nunc updating rule that satisfies these criteria? I start by
reviewing the distinction between updating and revision, an established distinction in the
literature of qualitative belief change.4
3 Here, I am not using "coherent" in the technical sense that a rational agent's credence should not lead her
to a Dutch book or money pump. Rather, I am using that word to mean freedom from incoherence, where
incoherence is defined to be logical or conceptual inconsistencies. Bricker (ms.) says: “One way for a
theory to be internally coherent is for it to be logically inconsistent, but I suppose there are other ways.
Moreover, if a theory is unfaithful to the notions it aims to elucidate, be they notions of ordinary or
scientific thought, that too is a form of incoherence.” (Bricker ms., p. 1.)
6666
Revision An agent revises her beliefs because she has acquired better information
about what the world is like.
Updating An agent updates her beliefs because she has noticed that the world has
changed.
For instance, compare the following ways in which Jane comes to know that Barack
Obama won the presidential election in 2008: First, it is 2009 and Obama has been
president for some time, but Jane has come to know this fact today. Second, it is the
morning after the day of the 2008 election, and Jane has come to know that he became
the winner of the election last night. The former example is a case of revision, while the
latter is a case of updating.
To see why this distinction is important, first look at this formulation of Strict
Conditionalization: for any proposition X,
(1) Ct(X)=Ct′(X/E)
where Ct and Ct′ are the agent’s credence functions at t and t′<t, and E is the totality of
her observation made during (t′,t]. Now, note that (1) entails:
(2) Ct(X/E)=Ct′(X/E)
4 For a general theory of qualitative belief change (called “the AGM model"), see Gardenfors et. al. (1985).
For the distinction made here, see Katsuno and Mendelzon (1992).
7777
where X and E are genuine propositions (or proposition-like entities whose truth-values
are fixed). This means that if an agent obeys Strict Conditionalization, she comes to
preserve her past conditional credences. As Christensen points out (2000), this is a form
of epistemic conservatism:
The reasonableness of attractive instances of conditionalization seems to flow directly from the
reasonableness of maintaining the relevant conditional degrees of belief. And these conditional
degrees of belief are valuable because they reflect past learning experiences. (Christensen 2000, p.
354)
In this sense, epistemic conservatism is a good thing because one has worked hard to
acquire valuable information about the world and to incorporate such information into
one’s belief state in the form of conditional credence.
But what if the world changes? Note that from an agent’s present point of view,
her past credence in X given Ei is the measurement of how probable it was that X was true
then given that Ei was true then. In this sense, her past conditional credences are outdated.
Consequently, epistemic conservatism is a disaster in this case: although the agent’s past
conditional credence in X given E was her degree of belief in X's then truth given E's then
truth, she came to preserve it as her degree of belief in X's truth now given E's truth now.
This is clearly unreasonable unless the agent has a reason to believe that the world is
likely to stay in the same state as before. I see no reason to have such a belief about the
world.
8888
Note that this was not a problem in the traditional version of Strict
Conditionalization, because that rule concerns only de dicto credences and evidence—the
degrees of belief in propositions with fixed truth-values and in evidence whose content’s
truth-value is similarly fixed. Still, the following variant of Strict Conditionalization
appears to be incorrect: for any tensed proposition X,
(3) Ct(X)=Ct′(X/E)
where Ct′ and Ct are B's credence functions at t and t′<t, and E is the totality of her
observations during (t′,t]. (3) implies
(4) Ct(X/E)=Ct′(X/E)
where X and E are tensed propositions. (4) is epistemologically dangerous for the reason
explained above, which suggests that although Strict Conditionalization is a proper rule
for the revision of de dicto credences, it is not a proper rule for the updating of de nunc
credences.
Think about this: In the past, you had conditional credences about what would
happen in a future time v given that you would observe E in v, and, from your present
point of view, it might be v now.5 Given this fact, if you know your present temporal
location, then the following principle seems to capture a correct form of epistemic
5 I am using “v” instead of “t” because I want to emphasize that the mentioned time is an inter“v”alized
time, rather than a momentary one.
9999
conservatism: Suppose that the agent B observes nothing during (t′,t), for some moments
t and t′<t. Then, for any proposition X,
(5) Ct(X)=Ct′(X is true in v/E is true in v)
where Ct′ and Ct are the agent’s credence functions at t and t′, E is the totality of the
agent’s observations made at t, and v is a temporal interval that B fully believes at t that
she is in.6
I believe that (5) is often a correct rule for updating de nunc credences. For if an
agent changes her de nunc credences obeying (5), then her resulting conditional credence
in X given E will be equal to her previous conditional credence in X’s truth in v given E’s
truth in v, where v is her present temporal location. For example, suppose that on
Tuesday Jake learns that there will be a form of precipitation today and sets his credence
in raining today to be equal to his conditional credence on Monday that it rains on
Tuesday given that there is a form of precipitation on Tuesday. In a sense, he preserves
his conditional judgment of how likely it is to rain on a day d given that there is a form of
precipitation on d, where d is the same day referred to on Monday as “Tuesday” and
referred to on Tuesday as “today.” In this case, it is intuitive that Jake has to update in
accordance with (5) because it is the best way to respect his past learning.
However, (5) is not general enough for our purpose, because it does not apply to
a case of temporal uncertainty such as the SB problem. We need another candidate for the
6 By “X is true in v,” I mean that X is true at any moment in interval v. Likewise for “E is true in v.” Plus, I
assume that interval v is sufficiently narrow, but I do not try to provide a criterion for sufficient narrowness
here.
10101010
general rule for de nunc updating. Consider this one: Suppose that an agent B observes
nothing during (t′,t) for some momentary times t, t′ such that t′<t. Then, for any tensed
proposition X,
(6) Ct(X)=Σi∈JCt′(X is true in vj/E is true in vj)Ct(it is vj),
where Ct′ and Ct are B's credence functions at t and t′<t, E is the totality of the de nunc
content of the observation made at t, and vjj∈J is a partition of temporal intervals such
that B fully believes at t that she is in one of vjj∈J.
To me, (6) is a plausible generalization of (5). To see why, let vj1≤j≤n be a
partition of temporal intervals each of which B thinks at t to be possibly her then temporal
location. By (5),
Ct(X) would be
Cn(X is true in v1/E is true in v1) if B were sure at t that it is v1,
Cn(X is true in v2/E is true in v2) if B were sure at t that it is v2,
...
Cn(X is true in vn/E is true in vn) if B were sure at t that it is vn,
Since B does not know what time it is, it appears to be natural to take the weighted
average of the above values, with the weights coming from B's credence at t that it is vj.
Thus, (6).
Equipped with (6), we are ready to answer the first question of the SB problem:
Let s be SB's last conscious moment on Sunday, m be the moment of wakeup on Monday,
11111111
and m+ be one minute after m when SB is told that it is Monday. Clearly, SB does not
make any observation during (s,m). By (6),
(7) Cm(H)=
Cs(H is true on Monday/W is true on Monday)Cm(it is Monday)+
Cs(H is true on Tuesday/W is true on Tuesday)Cm(it is Tuesday).
Since H is a genuine proposition and its truth-value is insensitive to time,
(8) Cm(H)=
Cs(H/W is true on Monday)Cm(it is Monday)+
Cs(H/W is true on Tuesday)Cm(it is Tuesday).
Since SB fully knew on Sunday that if she wakes up on Tuesday, then H is false,
(9) Cm(H)=Cs(H/W is true on Monday)Cm(it is Monday).
Since she fully expected on Sunday to wake up on Monday,
(10) Cm(H)=Cs(H)Cm(it is Monday).
Since her credence at s in H was 1/2 and she cannot be sure at m that it is Monday,
(11) Cm(H)=1/2Cm(it is Monday)<1/2.
12121212
This is a result favorable to the Thirder view and incompatible with the Halfer view.
Thus, (6) is not only a plausible principle for updating de nunc credences, but
also it provides the SB problem with a solution that is largely favorable to the Thirder
view. This means that it satisfies one of my criteria for an acceptable principle for de
nunc updating.7
However, there are two reasons to suspect that (6) is not the end of the story.
First, it follows from (6) that SB’s credence in H does not change from the moment of
wakeup on Monday to that of being told that it is Monday, but we have a reason to
consider this to be an incoherent result. To see the reason, first think about the following
instance of (5) (which is a special case of (6)): At m+, SB knows that it is Monday. Thus,
(12) Cm+(H)=Cm(H is true on Monday/MON is true on Monday),
where MON is the tensed proposition that it is Monday. Since H has a fixed truth-value
and she knows at m that MON is, of course, true on Monday,
(13) Cm+(H)=Cm(H)<1/2.
This result does not cohere with
7 Of course, this result may not be satisfactory to the Halfers. But I am focusing here on the facts that (i) the
provided solution is attractive to a large group of philosophers and (ii) it comes with an argument that
might be plausible even to some philosophers who initially opposed its conclusion.
13131313
(14) Cm(H/MON)=1/2,
which is provable from the very (6).8
To understand why, think about this matter from SB's point of view when she is
told that it is Monday. Previously, her credence in H was 1/2 given MON, and, by
learning that it is Monday now, she also learns that it was previously Monday. Intuitively,
her credence in H must increase back to 1/2 by this learning.
Second, (6) will not apply to a credal transition from t′ to t if the agent makes any
observation during (t′,t). For instance, SB experiences W when she wakes up on Monday;
thus, she makes a seemingly important observation between s (=the night on Sunday) and
m+ (=one minute after her wakeup on Monday). Although we can apply (6) to her credal
transition from s to m and to her credal transition from m to m+ (the second application is
seen in the problem discussed in the last paragraph), it would be better if we could
calculate SB's credence at m+ in H all at once from her credence distribution on Sunday
night. In this sense, (6) does not satisfy the criterion of versatility.
In sum, while I have a promising prototype for the general rule for de nunc
updating, it does not perfectly satisfy the three aforementioned criteria. My goal in this
dissertation is to find a general rule for de nunc updating which is fully versatile and
coherent, and which provides a fully intuitive solution for the SB problem.
D. Contents
This dissertation consists of six chapters:
8 Hint: Show first that Cm(H&MON)=Cm(T&MON). Since Cm(MON)=Cm(H&MON)+ Cm(T&MON),
Cm(H/MON)=Cm(H&MON)/Cm(MON)=Cm(H&MON)/2Cm(H&MON)=1/2. (See Chapter II for a full proof.)
14141414
Chapter I. Introduction
Chapter II. Updating with a Single Observation
Chapter III. Updating with a Sequence of Observations
Chapter IV. Updating with De Priori Information
Chapter V. Satisfaction of Desiderata
Chapter VI. Conclusion
The titles of Chapters I and VI are self-explanatory. Roughly, the main body of the
dissertation consists of three parts, each devoted to one of my criteria: (i) In Chapter II, I
discuss a relatively simple principle for de nunc updating, to solve the SB problem. (ii) In
Chapters III and IV, I generalize that simple principle into more versatile principles. (iii)
In Chapter V, I prove that the most general principle has several properties that we can
regard as forms of coherence. I provide more details below.
In Chapter II, I will discuss how a rational agent changes her credence in a
tensed proposition from t to t′, assuming that she receives no evidence during (t′,t). I will
present and defend the following principle for updating de nunc credences, which I call
“Shifted Jeffrey Conditionalization" or “SJC": Let X be any tensed proposition. Then,
roughly,
(15) Ct(X)=Σi∈I,j∈JCt′(X is true in vj/Ej is true in vj)Ct(Ej is true&it is vj),
15151515
where Ct′ and Ct are B's credence functions at t and t′, Eii∈I is a partition whose member
represents an observation she might be making at t, and vjj∈J is a partition whose
member represents a temporal interval that she might be in at t.9
Since (6) is a special case of SJC (or (15)), SJC suffers from the two problems I
discussed earlier. Figure 1 is a diagram showing how SB's credence in H changes in
accordance with SJC:
Figure 1: Update in Accordance with SJC. Not as versatile as wanted. Some incoherent
result.
where x is some value less than 1/2. First, SJC does not apply to the all-at-once updating
from s to m, which makes it not as versatile as we want. Second, SJC yields the counter-
intuitive result that her credence in H does not increase back to 1/2.
In Chapter III, I discuss how an agent can change her credence in a tensed
proposition from t′ to t, assuming that she makes a finite sequence of observations during
(t,t]. I will present a rule governing this type of updating, which I call “Sequential Shifted
Jeffrey Conditionalization” or “SSJC.” I cannot provide a proper formulation of this rule
here; simply, we do not have the necessary formal and conceptual resources yet. Instead,
9 Note that (15) is more general than (6) in that (15) incorporates uncertainty about what observation was
made as well as uncertainty about what time it is.
16161616
I provide its instance involving how SB's credence in H changes from the night on
Sunday (=s) to when she is told that it is Monday (=m+):
(16)
×
=
+
+
Monday isit and trueis presently, (d)
&Monday it was and true was ,previously (c)
(
Mondayon trueis (b)
&Mondayon trueis (a)
Monday/ on trueis (
)(
MON
W
C
MON
W
HC
HC
ms
m.
SB makes an observation twice during (s,m], the first time W and the second time MON.
In other words, she makes a sequence of observations <W,MON> during (s,m]. Here is
the core idea of (16): To find SB’s rational credence in H given this sequence of
observations, we need to specify, for each element of this sequence, the time of her
observing it, as you see in (c) and (d), and figure out her prior credence in H given that
each element of the sequence is true at those specified times, as you see in (a) and (b).
Hopefully, the reader will see how to generalize this idea into a formal rule for updating.
Unfortunately, SSJC is inconsistent. Figure 2 is a diagram showing how SB's credence in
H changes in accordance with SSJC:
Figure 2: Update in Accordance with SSJC. Versatile but inconsistent because x≠1/2.
where x is some value less than 1/2. As you see above, SSJC provides a different result
depending upon whether we apply it to the transition from s to m+ or to the transition
17171717
from s to m and then to that from m to m+. This makes it an unacceptable rule (unless
restricted by a suitable proviso).
In Chapter IV, I discuss how a rational agent changes her credence in a tensed
proposition from t′ to t, assuming that her observation may include information about her
temporal location at t′ or an earlier moment. The updating rule presented in this chapter
will be the most general updating rule discussed in this dissertation. I will call this rule
“General Shifted Jeffrey Conditionalization” or “GSJC.” Again, I do not try to present
the rule here. Instead, I discuss how it solves the problem that SB's credence in H does
not change at m+ although her then evidence MON is intuitively relevant to H. Under
several assumptions, we can derive the following claim from GSJC:
(17)
×
=
+
+
Monday it was ,previously (d)
&Monday isit and trueis presently, (c)
(
Monday isit presently, (b)
&Monday on trueis (a)
Monday/ on trueis (
)( MON
C
MON
HC
HC
mm
m.
In this updating process, MON is the only thing that SB has learned during (m,m+]. In a
trivial sense, we can say that <MON> is the sequence of observation that she has made
during (m,m+].
Here is the core idea in (17): To find her rational credence in H given this
sequence of an observation, we need to specify the time of her observing W and figure
out her prior credence in H given W’s truth at the specified time (which happens to be
Monday), as you see in (a) and (c). But that’s not all. We also need to specify what time it
was before she observed W and figure out her prior credence in H given W’s truth on
18181818
Monday plus the prior time’s being Monday, as you see in (b) and (d). I will call such
information about what time it was at the prior time “de priori information.”10
Let us see how this modification provides a better model for how SB's credence
in H changes. According to GSJC, SB’s credence in H changes as in Figure 3:
Figure 3: Update in accordance with GSJC. Versatile, coherent, providing results
compliant with the popular Thirder view.
where x is some value less than 1/2. As Figure 3 demonstrates, GSJC provides a coherent
and intuitive model for the change of SB's credence in H. This model also complies with
the view of Thirders, regarding both questions asked earlier.
In Chapter V, I argue that GSJC has several properties desirable for any cogent
rule for updating. In particular, I will show that (i) GSJC can be regarded as a binary
relation between the given agent’s credence functions (at different times) and (ii) as such,
GSJC is transitive, if all of the agent’s credence functions are synchronically coherent.
See Figure 4:
10 I owe this term to Gareth Matthews.
19191919
Figure 4: Transitivity of GSJC. Here, I(t,t′) is information observed during (t,t′].
Put another way, GSJC provides the same result whether you update all at once from t1 to
tn or step-by-step from t1 to t2, t2 to t3, … to tn.
In Chapter VI, I discuss (i) how to modify GSJC into a general rule for de se
updating, (ii) how to overcome a potential problem of GSJC, and (iii) whether there is
any comparably plausible but simpler rule for updating. I conclude that GSJC is likely to
be identical or very close to the general rule for de se updating.
Through these discussions, I will defend my view that GSJC is the rational rule
for updating de nunc credences. I will argue that GSJC provides not only an ideal
solution for the Sleeping Beauty problem but also a versatile and coherent general rule
for de nunc updating.
20202020
CHAPTER II
UPDATING WITH A SINGLE OBSERVATION11
A. Introduction
SB problem 0. Suppose that Sleeping Beauty (hereafter: SB), a paragon of probabilistic
rationality, knows the following facts on Sunday: A group of evil experimenters will put
her to sleep on that day. Next, they will toss a fair coin. Case 1: (H) The coin lands heads.
In this case, the experimenters will wake SB only on Monday. Case 2: (T) The coin lands
tails. In this case, they will wake SB for the first time on Monday, inject her with a drug
that erases her memory of Monday, and then wake her for the second time on Tuesday. In
either case, the experiment is over on Wednesday.
For brevity, let s be the last moment on Sunday at which SB is conscious and let
m be the moment of waking up on Monday. Accordingly, let Cm and Cs be her credence
functions at m and s. The question is: “What is SB’s credence at m in H?” There have
been two dominant answers: According to the Thirder view, Cm(H)=1/3 (Elga 2000).
According to the Halfer view, Cm(H)=1/2 (Lewis 2001). Both views have good
arguments in their favor.
Halfers contend: Let W be that SB wakes up today with the memory of Sunday
as the last memory. When SB was put to sleep on Sunday, she fully expected to receive
W as her next evidence. Furthermore, that she has awakened today with such and such
11 This chapter is identical to my paper published in Synthese (Kim 2009) except for several changes of
notation and correction of typos.
21212121
memory seems irrelevant to whether the coin lands heads or tails. Hence, W is neither
new nor relevant to H. But the thesis below derives from the standard rule for credence
updating:
(1) If no new evidence relevant to X is received,
then it is irrational for an agent to change her credence in X.
Thus, SB doesn’t change her credence in H after waking up on Monday. It is not
controversial that SB assigns the credence of 1/2 to H on Sunday night. Therefore, she
assigns 1/2 to H after waking up on Monday. (Lewis 2001, 174.)
Thirders argue: Let MON be that it is Monday, and TUE be that it is Tuesday.
Then, we can define H1, T1, and T2 as below:
H1: H&MON
T1: T&MON
T2: T&TUE
Obviously, these exhaust the possibilities open to SB when she wakes up on Monday. On
the one hand, suppose that SB was immediately told that it’s Monday after waking up on
Monday. In this scenario, she would assign the same credence of 1/2 to H and T. Hence,
Cm(H/MON)=1/2=Cm(T/MON). It follows that
(2) Cm(H1)=Cm(T1).
22222222
On the other hand, assume that SB was told immediately after waking up on Monday that
the coin landed on tails. The evidence she receives on Monday, in this scenario, is
compatible with either MON or TUE; for, if the coin lands tails, she wakes up both on
Monday and Tuesday. By a principle of indifference, it seems rational to assign the same
credence to MON and TUE, given T; formally, Cm(MON/T)=Cm(TUE/T). It follows that
(3) Cm(T1)=Cm(T2).
In sum, Cm(H1)=Cm(T1)=Cm(T2)=1/3. But H1 is the only possibility in which the coin
lands heads. Hence, Cm(H)=1/3. (Elga 2000, 143-144.)
Which side made a mistake? Theses (1) and (2) contradict each other.12
Hence,
Halfers, who accept (1), are bound to reject (2), and Thirders, who accept (2), are
committed to the rejection of (1). Thus, each side complains that the other’s argument is
factually incorrect. The problem is that neither side has been able to explain why the
other side’s key premise is wrong. For this reason, the debate has continued. I suspect,
however, that both sides have missed an important point. We know quite well SB’s belief
state on Sunday night; on that night, her credence in H was 1/2, and she knew how the
experiment would proceed. W is the only evidence she acquires until she wakes up next
morning. Isn’t it then simply a matter of applying Strict Conditionalization (hereafter:
SC), the traditional principle for updating credences, to SB’s credal transition from s to m?
12 Suppose theses (1) and (2). Hence, Cm(H)=Cs(H)=1/2 and Cm(H1)=Cm(T1). Unless Cm(T2)=0,
Cm(H1)<Cm(T1)+Cm(T2). Since H1 exhausts the H-possibility and T1 and T2 exhaust the T-possibilities,
Cm(H)<Cm(T). Therefore, 1/2=Cm(H)<1/2. Done.
23232323
Unfortunately, the problem is not that simple. It is an instance of SC that
Cm(H)=Cs(H/W). However, SB fully knew on Sunday night that she did not wake up on
that day with the memory of Sunday as the last memory, and so Cs(H/W) is undefined.
Nevertheless, I believe that there exists an updating principle applicable to SB’s
credal transition from s to m. My first goal in this chapter is to find an updating rule for
when the domains of an agent’s credence functions include tensed propositions or
proposition-like entities whose truth-values are possibly different depending upon the
time of evaluation. I will call this new updating principle “Shifted Jeffrey
Conditionalization” (hereafter: SJC). I will argue that because W is a tensed proposition,
SB has to use SJC, not SC, for her credal transition from s to m.
My second goal in this chapter is to explore the ramifications of this new
updating rule concerning the SB problem. I shall make three claims: First, thesis (1) is
disprovable under SJC. Second, thesis (2) is provable under SJC. Third, thesis (3) is
neither provable nor disprovable in any obvious way.
If these claims are true, who wins in the Sleeping Beauty debate? On the one
hand, Halfers are clearly not the winners. For their argument is unsound if thesis (1) is
false and SB’s credence at m in H is less than 1/2 if thesis (2) is true. On the other hand,
this does not necessarily mean a victory for the Thirders. For thesis (3) is an essential
element of their view but SJC does not obviously support it.
Consequently, I partially accept the Thirder view but take a more lenient position:
I will argue that SB’s credence in H is less than 1/2 when she wakes up, but I will remain
silent about what the value should be. Call this “the Lesser view.” In this chapter, I will
defend it.
24242424
I will proceed in the following order: In Section B, I will clarify my assumptions
and terminology and review traditional updating principles. In Section C, I will argue that
those updating principles do not work for beliefs and evidence whose truth-values are
different relative to time and/or individual. In Section D, I will present an alternative
updating principle, SJC, for such beliefs and evidence. In Section E, I will defend SJC by
extending Gaifman’s influential view of expert principles. Finally, in Section F, I will
apply SJC to the SB problem. As a result, the Halfers’ thesis (1) will be criticized and the
Thirders’ thesis (2) will be defended.
B. Background
In this section, I will clarify my assumptions and terminology about beliefs, contents, and
credences.
First, belief is a relation between an agent and a proposition-like entity. For
example, consider Jane’s belief that (C) Caesar crossed the Rubicon in 49 BC. According
to my assumption, this belief is a relation between Jane and C.
Second, the truth-values of some beliefs remain the same whoever has them or
whenever they are had. Jane’s belief in the above paragraph is a good example. Let’s call
such a belief a “de dicto belief” and its content a “(genuine) proposition.” I consider such
a belief to be purely about which possible world the agent is located in.
Third, the truth-values of some beliefs are different relative to times and/or
individuals. For example, consider Jane’s belief expressed by “I am 15 years old.” The
content of this belief will be true of anyone who is 15 years old, but won’t be true of
anybody who is younger or older. Let’s call such a belief an “irreducibly de se belief”
25252525
and its content an “irreducibly centered proposition.” I consider a belief of this type to be
at least partially about what time it is and/or who the agent is.
Fourth, I will call any de dicto or irreducibly de se belief simply “a de se belief”
and its content “a centered proposition.”13
Fifth, belief is not all-or-nothing but comes in degrees. Degrees of belief, called
“credences,” are probabilities in that they satisfy Kolmogorov’s three axioms: Non-
Negativity, Normality, and Additivity.
Sixth, I will call the degree of a de dicto belief a “de dicto credence,” that of an
irreducibly de se belief an “irreducibly de se credence,” and that of a de se belief a “de se
credence.”
So far, I have clarified my assumptions and terminology. Now, to the question:
“What is the correct rule for updating de se credences?” To answer, it is a good idea to
review the traditional rules for updating de dicto credences.
First, how is a rational agent supposed to update her de dicto credences given
certain evidence? Consider the strongest proposition E such that a rational agent B
becomes certain at tn+1 of E, as a result of her experience at tn+1.14
Then, B should update
by Strict Conditionalization: (SC) for any proposition X, Cn+1(X)=Cn(X/E)=df
Cn(X&E)/Cn(E), where Cn(E)≥0. To see how SC works, consider this example: Example
13 I am following Lewis (1979) in defining de se belief in this way: “I say that all belief is ‘self-locating
belief.’ Belief de dicto is self-locating belief with respect to a logical space; belief irreducibly de se is self-
locating belief at least partly with respect to ordinary time and space, or with respect to the population”
(Lewis 1979, 522).
14
For simplicity, I will assume that the given agent makes observations at only a countable number of
moments in her life. Let’s call them “epistemic moments.” From now on, I will use “tα” to refer to a series
of epistemic moments, where α indicates the order and contiguity of those moments. Hence, for any
m,n∈N, tm is later than tn iff m>n, and for any n∈N, tn+1 is the epistemic moment next to tn. In addition, I
will use “Cα” to refer to the given agent’s credence function at tα.
26262626
1. Jane’s previous conditional credence in a coin’s landing heads was 3/4 given that it is
tossed. After she receives the evidence that it was tossed, her present unconditional
credence in the coin’s landing heads becomes 3/4.
Second, how does a rational agent update her de dicto credences if no
proposition meets the condition satisfied by E in the last paragraph? Suppose that Eii∈I
is a partition such that for any i∈I, an agent B’s credence in Ei is directly set by her
experience at tn+1. I will call each Ei “(B’s) observation proposition at tn+1.” Let Eii∈J be
a subset of this partition (so J⊆I) such that Cn+1(Ei)>0 for any i∈J. If also Cn(Ei)>0 for
any i∈J, then we call Eii∈J “(B’s) observation partition at tn+1.” In such a case, Richard
Jeffrey suggests that B should update her de dicto credences by Jeffrey Conditionalization:
(JC) for any proposition X, ∑∈
++ =Ji
ininn ECEXCXC )()/()( 11 (Jeffrey 1990, pp. 164-83).
To see how JC works, consider this example: Example 2. At 2:00 PM, Jane is looking at
a piece of vegetable under a dim light, uncertain whether it is green or violet. Hence, she
is uncertain about which of G (“this piece of vegetable is green”) and V (“this piece of
vegetable is violet”) is true. Still, her experience somehow influences her credences in G
and V; consequently, her credences at 2:00 PM in G and V are 0.3 and 0.7. Then, what
should her credence be in C (“it is a piece of cabbage”)? Her conditional credence at 1:59
PM in C was 0.6 given G and 0.2 given V. By JC, C2:00 PM(C)=C1:59 PM(C/G)C2:00 PM(G)+
C1:59 PM(C/V)C2:00 PM(V)=0.32.
If, as I have demonstrated, SC or JC is the rule for updating de dicto credences,
what then is the rule for updating de se credences? According to David Lewis, we can
easily find a candidate for such a rule: just replace “de dicto” with “de se” and
27272727
“proposition” with “centered proposition” in SC (Lewis 1979, 534). Another candidate
can be found by carrying out the same replacement in JC.
However, I believe that the de se versions of SC and JC are incorrect. For they
have a common problem, which I discuss in the next section.
C. A Problem of the De Se Versions of SC and JC
I want to show that the de se versions of both SC and JC are untenable. However,
criticizing them will be a tedious job if done one by one. A more efficient method will be
to first find a thesis common to the two principles and, second, show that this common
thesis has a fatal problem. Is there such a thesis?
According to Richard Jeffrey (1984, p. 135), we can easily prove this:
(C) Suppose that both Cn and Cn+1 satisfy Kolmogorov’s axioms. Let E be the
agent’s total evidence at tn+1. Then, (a) for any proposition X, Cn+1(X)=Cn(X/E) iff
(b) Cn+1(E)=1 and (c) for any proposition X, Cn+1(X/E)=Cn(X/E).
In other words, an agent is a strict conditionalizer iff she certainly believes her total
evidence, and its probabilistic relevance to any other belief is unchanged by updating.
Following Jeffrey’s terminology, let’s call condition (c) “Rigidity.” We can also prove
this broader claim (Jeffrey 1984, p. 136):
(K) Suppose that both Cn and Cn+1 satisfy Kolmogorov’s axioms. Let Eii∈I
be the agent’s observation partition at tn+1. Then, (d) for any proposition X,
28282828
∑∈
++ =Ii
ininn ECEXCXC )()/()( 11 iff (e) for any proposition X and i∈I,
Cn+1(X/Ei)=Cn(X/Ei).
In other words, an agent is a Jeffrey conditionalizer iff her old and new credence
functions satisfy Rigidity. According to (C) and (K), the truth of Rigidity regarding each
member of her observation partition is a common necessary condition of SC and JC.15
Now, replace “proposition” with “centered proposition” in (C) and (K); still, they are
provable claims. Let’s call the results of this replacement within (a), (d), and (c)/(e) “the
de se versions of SC, JC, and Rigidity.” Then, the de se version of Rigidity is the
common necessary condition of those of SC and JC.
But Rigidity has a fatal flaw. It conflicts with the logic of de se beliefs,16
in that a
probabilistic updating pattern it supports often leads to deductively invalid reasoning.
Think about this example: Example 3. Let R be the centered proposition expressed by “it
is raining now,” and P be the one expressed by “some form of precipitation is occurring
now.” Assume that P, not-P is Jake’s observation partition at 2:00 PM. For our purpose,
it is best to discuss the present example from the first-person point of view; hence, we let
Cprev be Jake’s credence function at 1:59 PM and Cnow be his credence function at 2:00
15 If the agent has certain total evidence E, then E is the sole member of her observation partition.
16
Meacham (2008) provides a simpler argument against the de se version of SC: Once a de se SC-er
becomes certain that it is 9:00 AM, she cannot abandon that belief at 9:01 AM. Since such abandonment
occurs too often, the de se version of SC cannot be true. This argument seems to be sound, but I have two
reasons to look for an alternative. First, Meacham’s argument does not show the incorrectness of the de se
version of JC. Second, his argument cannot defeat the de se version of SC formulated in terms of primitive
conditional credences. For the problem raised by Meacham is a form of the zero-denominator problem,
which can be avoided if we define unconditional credence in terms of conditional credence rather than
doing the opposite (Hajek 2003).
29292929
PM. Additionally, assume that Cprev(P)=0.5 and Cnow(P)>0. Obviously, we can derive this
fact from the de se version of Rigidity:
(4) If Cprev(R/P)≈1, Cnow(R/P)≈1.17
From the suppositions, we can easily show that (i) if Cprev(P⊃R)≈1, then Cprev(R/P)≈1 and
(ii) if Cnow(R/P)≈1, then Cnow(P⊃R)≈1.18
Hence, this is true:
(5) If Cprev(P⊃R)≈1, then Cnow(P⊃R) ≈1.
This means that the following normative sentence is true of Jake at 2:00 PM:
(6) If I strongly believed P⊃R previously,
I must strongly believe P⊃R now.
Practically, following (6) amounts to doing this type of reasoning:
(7) It was previously the case that, if P, then R.
Therefore, it is now the case that, if P, then R.
17 Here, “≈” is a synonym of “is almost identical to.”
18
Here, “⊃” is the material implication connective. To prove (i), suppose that Cprev(P⊃R)≈1. We know that
Cprev(P⊃R)=Cprev(not-(P¬-R))=1-(1-Cprev(R/P))Cprev(P). By the assumption that Cprev(P)=0.5,
Cprev(P⊃R)=1/2(1+Cprev(R/P)). Thus, Cprev(R/P)=2Cprev(P⊃R)-1. By supposition, Cprev(R/P) ≈1. To prove
(ii), suppose that Cnow(R/P) ≈1. Since Cnow(P⊃R)=1-(1-Cnow(R/P))Cnow(P), Cnow(P⊃R)≈1. Done.
30303030
However, this argument form is invalid. For it is surely possible that the premise is true
but the conclusion is false: It was previously raining, and it is now snowing. Hence, the
de se version of Rigidity may lead to invalid reasoning.19
This is a good reason to
abandon it.
Since the de se version of Rigidity is a common necessary condition of SC and
JC, we should reject both.
D. Shifted Jeffrey Conditionalization
If neither SC nor JC is an acceptable principle for de se updating, then what is? I do not
have a general principle for updating de se credences yet. In this section, however, I will
suggest a new principle for updating the degrees of some special de se beliefs.
Before presenting this new updating principle, I need new notions related to
beliefs, contents, credences, and time:
First, there are de se beliefs whose truth-values are different depending upon
when they are had but not upon who has them. For instance, think about Jane and Jack’s
common belief that it rains today in Boston. It can have different truth-values at different
times: It is possible that this belief is true on Monday but false on Tuesday. However, it
cannot have different truth-values for different people at any given time: It is impossible
that this belief is true for Jane but false for Jack on any given day. Let’s call a de se belief
of this type an “irreducibly de nunc belief” and the content of an irreducibly de nunc
19 Admittedly, this result is dependent upon the suppositions that (i) Cprev(P)=0.5 and (ii) Cnow(P)>0.
However, neither supposition includes anything that possibly justifies the reasoning of (7). For (i) merely
means that Jake is neutral between P and not-P and (ii) just means that he doesn’t rule out P.
31313131
belief an “irreducibly tensed proposition.” I consider a belief of this type to be partially
about what time it is.
Second, we will call a de dicto or irreducibly de nunc belief a “de nunc belief,”
and we will call a genuine or irreducibly tensed proposition a “tensed proposition.”20
Third, we will call the degree of an irreducibly de nunc belief an “irreducibly de
nunc credence” and that of a de nunc belief a “de nunc credence.”
Fourth, I introduce this definition: For any tensed proposition X and (temporal)
interval v,
The truth-value of X is invariant within v iff for any moments t and t′ in v, X’s
truth at t logically implies X’s truth at t′.
For instance, consider tensed proposition R expressed by “it rains in Boston at some time
today.” The truth-value of R is invariant within Monday; for, if R is true/false at some
moment on Monday, then R is true/false at any other moment on Monday. Similarly, the
truth-value of R is invariant within Tuesday. However, R can have different truth-values
on Monday and on Tuesday.
In addition to these notions, we need an adequate formal language to formulate a
new updating principle. Hence, I construct such a language L:
First, we need the traditional probability language’s logical connectives,
arithmetic operators, and probability function letters in our new language.
20 For the deductive theories of tensed propositions, see Prior (2003) or Rescher and Urquhart (1971).
32323232
Second, as we need propositional letters in the traditional probability theory, we
need tensed propositional letters in our new language. For this, we reserve “E,” “F,” “X,”
“Y,” and “Z.”
Third, we use “t” as a moment letter, which denotes a moment or a point-sized
temporal location, and “v” as an interval letter, which denotes an interval or a continuous
class of moments. Together, we call them time letters.
Fourth, given a time letter, we use the corresponding capital letter as a special
tensed proposition letter, denoting the tensed proposition that the present moment is or
belongs to the time denoted by the variable. For example, if “t” is the letter denoting the
moment of 9:00 AM on Sep. 4th
in 2006, then “T” is the letter denoting the tensed
proposition that the present time is exactly that moment. Similarly, if “v” is the interval
letter denoting the day of Sep. 4th
in 2006, then “V” is the letter denoting the tensed
proposition that it is that day.
Fifth, we introduce binary operators “at” and “in.” Let “ϕ,” “τ,” and “ν” be
schematic letters replaced with a tensed propositional letter, a moment letter, and an
interval letter, respectively. Here is a meaning schema for “at”:
“ϕ at τ” means that ϕ is true at τ.
For instance, let “R” mean the tensed proposition that it rains now in Boston and let “t”
denote the moment of 9:00 AM on July 18th
2006. Then, “R at t” denotes the proposition
that it rains in Boston at that moment. In addition, we introduce this abbreviation:
33333333
“ϕ in ν” abbreviates “(∀t∈ν)(ϕ at t).”
In other words, “ϕ in ν” means that ϕ is true throughout ν. For example, let v be July 18th
2006. Then, “R in v” means that it rains in Boston throughout July 18th
2006.
Now, we are ready to discuss a principle for updating de nunc credences. First,
we consider an agent who has a tensed proposition E as certain total evidence and fully
believes that she is located in v. For such an agent, I recommend this method of updating
her credence in a tensed proposition, which I call “Shifted Strict Conditionalization”:
(SSC) Cn+1(X)=Cn(X in v/E in v) if E is the agent’s certain total evidence at tn+1
and she fully believes at tn+1 that it is v,
where Ct(E in v)>0, and the truth-values of X and E are invariant within v. In other words,
if E is a rational agent’s certain total evidence at t+1, and she fully believes at tn+1 that it
is v, her credence at tn+1 in a tensed proposition X is the same as her previous conditional
credence in [X’s truth in v] given [E’s truth in v], as long as the conditional credence is
defined and neither X nor E can have different truth-values at any two moments in v.
In order to see how SSC works, consider this example: Example 4. Let P be that
some form of precipitation occurs today in Boston and R be that it rains today in Boston.
Suppose that on Sunday, Jane’s conditional credence in [R’s truth on Monday] given [P’s
truth on Monday] is 0.3.21
On Monday, she learns and becomes certain that some form of
precipitation occurs today in Boston, and today is Monday. What, then, is Jane’s rational
21 In this paper, I use brackets simply to avoid scope confusion.
34343434
credence on Monday in its raining today in Boston? My answer: CMON(R)=CSUN(R is true
on Monday/P is true on Monday)=0.3 by SSC.
Second, consider an agent B who possibly lacks certain total evidence or is
uncertain of what time it is. In order to capture such uncertainties, we assume two
partitions Eii∈I and vjj∈J, such that Eis are the tensed propositions, such that B’s
credences in Eis are directly set by her experience at tn+1, and such that vjs are temporal
intervals covering the minimal interval that B fully believes at tn+1 that she is located in.
Given these partitions, we will call Eis “(B’s) observation propositions at tn+1” and Vjs
“(B’s) temporal location propositions at tn+1.” Roughly, an agent lacks certain total
evidence iff she is uncertain of which of her observation propositions is true, and she is
uncertain of what time it is iff she is uncertain of which of her temporal location
propositions is true.
It’s possible to unify these two dimensions of uncertainty into one. Consider
Ei&Vj<i,j>∈I×J, consisting of consistent conjunctions of the members of the above two
partitions. Let’s call any such conjunction “(B’s) time-observation proposition at tn+1.” To
B, one of her time-observation propositions at tn+1 is a candidate for the true conjunction
of her observation and temporal location propositions at tn+1. Roughly, an agent lacks
certain total evidence and/or is not sure of what time it is iff she is uncertain of which of
her time-observation propositions is true.
Naturally, we focus upon B’s time-observation propositions whose truth she
doesn’t completely rule out at tn+1. Thus, let K⊆I×J be the class of <i,j>s such that
35353535
Cn+1(Ei&Vj)>0, where Cn+1 is B’s credence function at tn+1. If Cn(Ei at vj)>0 for any <i,j>
in K, I call Ei&Vj<i,j>∈K “(B’s) time-observation partition at tn+1.”
Finally, we formulate a method of updating credences in tensed propositions,
which I call “Shifted Jeffrey Conditionalization”:
(SJC) ∑>∈<
++ =Kji
jinjijnn VECvinEvinXCXC,
11 )&()/()( if E i& V j < i, j>∈ K
is the agent’s time-observation partition at tn+1,
where, for any i∈I and j∈J, the truth-values of X and Ei are invariant within vj. In other
words, if an agent lacks certain total evidence and/or is uncertain of what time it is at tn+1,
then her rational credence at tn+1 in a tensed proposition X is the weighted average of the
results of applying SSC to X with various time-observation propositions at tn+1, in which
the weights are her credences at tn+1 in the time-observation propositions, as long as
neither X nor any of Eis logically can have different truth-values at any two moments
within each interval vj.
To see how SJC works, think about the following examples: Example 5. Let “P”
and “R” have the same meanings as in Example 4. On Sunday, Jane’s conditional
credence in [R’s truth on Monday] given [P’s truth on Monday] is 0.5, and her
conditional credence in [R’s truth on Monday] given [not-P’s truth on Monday] is, of
course, 0. On Monday, some perfectly reliable person tells her that it is Monday and
something about today’s weather, but she doesn’t hear the latter information clearly. Has
he said that some form of precipitation occurs, or that it doesn’t occur, today in Boston?
She is unsure. Her credence in the former is 0.7 and that in the latter is 0.3. In this case,
36363636
what is Jane’s credence on Monday in its raining in Boston on that day? My answer:
CMON(R)=CSUN(R is true on Monday/P is true on Monday)CMON(P)+CSUN(R is true on
Monday/not-P is true on Monday)CMON(not-P)=0.35 by SJC.
Example 6. Again, keep the meanings of “P” and “R” the same. On Sunday,
Jane is put to sleep with one of two drugs. The first drug’s effect lasts for just one night,
making her wake up on Monday. The second drug’s effect lasts longer, making her wake
up on Tuesday. Her conditional credence on Sunday in [R’s truth on Monday] given [P’s
truth on Monday] is 0.8, and her conditional credence in [R’s truth on Tuesday] given
[P’s truth on Tuesday] is 0.2. On Monday, Jane is told by a perfectly reliable person that
some form of precipitation occurs today in Boston, and, not knowing which drug she took,
Jane assigns the credence of 0.4 to its being Monday and that of 0.6 to its being Tuesday.
In this case, what is Jane’s rational credence on Monday in R? My answer:
CMON(R)=CSUN(R is true on Monday/P is true on Monday)CMON(it is Monday)+CSUN(R is
true on Tuesday/P is true on Tuesday)CMON(it is Tuesday)=0.44 by SJC.
Despite the complicated appearance, SSC and SJC are actually quite intuitive.
Certainly given E as total evidence and v as her temporal location, an agent should assign
to X the previous conditional credence in X’s truth in v given E’s truth in v, not that in X
given E. For the latter conditional credence is the (previous) conditional credence in X’s
previous truth given E’s previous truth, which seems irrelevant to the (present) credence
in X’s present truth. Once accepting this, it seems natural that if the agent lacks certain
total evidence and/or does not know what time it is, the new credence in X should be the
weighted average of the previous conditional credence in [X’s truth in vj] given [Ei’s truth
37373737
in vj], where Eis and vjs are the candidates for the presently true observation proposition
and the temporal interval in which she is presently located.
In this section, I have presented a new principle for updating credences and
illustrated how that principle works with examples. The intuitive answers yielded by SJC
provide evidence in support of it, but additional argument is needed in order for us to
accept its validity.
E. Shifted Rigidity as a Conditional Expert Principle
In this section, first, I will present a principle entailing SJC (which has SSC as a special
case), and, second, I will discuss how that principle can be promoted by an intuitive
expansion of Gaifman’s Expert Principle (Gaifman 1988).
First, the principle: I call it “Shifted Rigidity.” Let Ei&Vj<i,j>∈K be the agent’s
time-observation partition at tn+1. Then, for any tensed proposition X and any <i, j> in K,
(SR) Cn+1(X/Ei&Vj)=Cn(X in vj/Ei in vj),
where the truth-values of Ei and X are invariant within each vj. From the agent’s point of
view, this says, “The present relevance of Ei to X, on the additional condition that I am
located in interval vj, is the same as the previous relevance of [Ei’s truth in vj] to [X’s
truth in vj].”
Obviously, SR entails SJC.22
This means that we can argue for the latter by
defending the former. But how can we defend SR? I believe that Gaifman’s discussion of
22 Let X be any tensed proposition and Ei&Vj<i,j>∈K⊆I×J be the agent’s time-observation partition at tn+1.
Suppose that Cn+1(X/Ei&Vj)=Cn(X in vj/Ei in vj) for any <i, j>∈K, where the truth-values of X and Ei are
38383838
Expert Principles provides a clue to this question. Here is the schema of his so-called
Expert Principles:
(Expert) C(X/pr(X)=r)=r, for all r such that C(pr(X)=r)>0.
Here, C is an agent’s credence function and pr is the agent’s “expert probability
function.” What is an expert probability function? Roughly speaking, it is a probability
function that, once it is known to the agent, will be adopted by her as her own credence
function. For instance, if you consider a local weather forecaster to be an expert for your
local weather, then C(rain/pr(rain)=r)=r where C is your credence function and pr is the
weather forecaster’s. Here, pr does not have to be a subjective probability function. When
pr is the objective chance function P, you get:
(Principal Principle) C0(X/P(X)=r)=r for any r such that C0(P(X)=r)>0,23
where C0 is an agent’s initial credence function. When pr is the agent’s future credence
function at the next epistemic moment, you get:
(Reflection) Cn(X/Cn+1(X)=r)=r for any r such that Cn(Cn+1(X)=r)>0.24
invariant within each vj for any i∈I and j∈J. Since the partition exhausts Ei&Vj such that Cn+1(Ei&Vj)>0,
Cn+1(∨<i,j>∈K (Ei&Vj))=1. Thus, Cn+1(X)=Cn+1(X&∨<i,j>∈K (Ei&Vj))=∑<i,j>∈KCn+1(X/Ei&Vj)Cn+1(Ei&Vj)=(by
supposition) ∑<i,j>∈K Cn(X in vj/Ei in vj)Cn+1(Ei&Vj). Done. (The other direction is also provable under a few
plausible assumptions but nothing in this paper hangs upon that direction of the equivalence. Still, the
equivalence is interesting because it is analogous with the equivalence between Rigidity and JC.) 23
The original version of PP is more general in that the condition can include additional information as
long as it’s admissible (Lewis 1987).
39393939
Why will an agent’s future credence function be her expert function? If the given agent is
not forgetful, she usually will be more knowledgeable in the future than in the present.
(This is a good way to see why the Reflection Principle fails in the counterexample of
Talbott (1991), in which the agent is forgetful.) By contrast, since an agent in the past is
typically less knowledgeable than the same agent in the present, an agent’s past credence
function can’t be her present expert function.
So far, so good. Now, consider this example.25
Example 7. An investor consults
a very trustworthy stock market expert. The problem is that the investor cannot reveal to
the expert some insider information she has that a company will release a new product
next month. The expert’s opinion is generally unfavorable to that company, but his
opinion conditioned upon the insider information is quite favorable. In that case, it will be
rational for the investor to make a judgment on the basis of the expert’s conditional (on
the insider information) opinion. In this situation, it seems to be rational for the investor
to adopt the expert’s conditional credence function on the product release information as
her own credence function. This intuition can be generalized as follows:
(Conditional Expert) C(X/E&pr(X/E)=r)=r, for all r such that
C(E&pr(X/E)=r)>0.
24 The original version of Reflection is more general in that the agent’s future credence function can be
from a farther future than tn+1 (van Fraassen 1984). 25
Hall discusses the same idea (Hall 2004, p. 100).
40404040
Here, C is an agent’s credence function and pr is the agent’s conditional expert
probability function on E. By “conditional expert probability function on E,” I mean a
probability function pr such that, once the function pr(-/E) and the truth of E are known
to the agent, the agent will adopt it as her credence function. Given this definition, the
Conditional Expert Principle is a natural expansion of the Expert Principle.
The Conditional Expert Principle provides a new way of understanding Rigidity.
I suggest a sub-principle of Conditional Expert below:
(Backward Reflection) Cn+1(X/Ei&Cn(X/Ei)=r)=r, for all r such that
Cn+1(Ei&Cn(X/Ei)=r)>0, where Ei is a member of the agent’s observation
partition at tn+1.
To the agent at tn+1, Ct is her previous credence function at tn. I suggest that it also must
be the agent’s conditional expert probability function at tn+1 on Ei. Why? An agent
usually has no choice but to depend on her previous credence to form the present one;
however, she is also aware that her previous credence distribution was built without her
present experience. Thus, an agent’s relation to her past credence distribution is similar to
the investor’s relation to the stock market expert in Example 7: Due to the informational
impoverishment of the agent’s previous self, it may be irrational that her present credence
in X is r given that her previous credence in X was r. Still, it is rational that her present
credence in X is r given that Ei is the true member of her present observation partition,
and her previous credence in X given Ei was r. This is because if Ei is true, then her
previous credence function conditional on Ei was a judgment made on the basis of all
41414141
information that she previously had plus the member of her observation partition actually
confirmed by her present experience.
Therefore, let’s make a plausible conjecture that Backward Reflection is usually
true of a rational agent. If we assume that the given agent correctly knows her credence
function and later remembers it with perfect confidence, Rigidity will obviously follow
from Backward Reflection.26
This provides a new way of understanding why we should
accept Rigidity.
The last step in our expansion of Gaifman’s principle is to apply the idea to
tensed propositions, especially concerning Backward Reflection. However, I suggest that
we need substantial modification to do so. Why? Consider this example. Example 8.
Again, let R be that it rains today in Boston and let P be that some form of precipitation
occurs today in Boston. In this example, an agent B at 9 AM on Monday (hereafter: tm),
knowing that it’s Monday, regards herself at 9 PM on Sunday (hereafter: ts) as an expert
about local weather in Boston except that she didn’t know whether there would be
precipitation on Monday. At tm, B learns that (i) P is true. At ts, (ii′) B’s credence in R
given P was 0.1, but (iii) her credence in [R’s truth on Monday] given [P’s truth on
Monday] was 0.3. I make two claims: First, it is not the case that B’s rational credence at
tm in R is 0.1 given (i) and (ii′), but, second, her rational credence at tm in R is 0.3 given (i)
and (iii).
26 Let r be Cn(X/E). If the agent remembers her past credence distribution with perfect confidence,
Cn+1(Cn(X/E)=r)=1. Thus, Cn+1(X/E)=Cn+1(X/E&Cn(X/E)=r)=(by Backward Reflection)r=Cn(X/E). Done.
42424242
To understand why, look at Figure 5:
Figure 5: Rain and Precipitation. The tensed proposition at the arrow tail is true on the
day of the tail’s column iff the event at the arrow head occurs on the day of the head’s
column.
On the one hand, B’s opinion, represented by the conditional credence at ts in R given P
seems to be relevant only to whether raining happens on Sunday given that a form of
precipitation occurs on Sunday. Hence, it is irrelevant to whether raining happens on
Monday given that a form of precipitation occurs on Monday, which B’s credence at tm in
R is factually about. This suggests that B’s rational credence at tm in R is not necessarily
0.1 given (i) and (ii′). On the other hand, B’s opinion at ts, represented by the conditional
credence at ts in [R’s truth on Monday] given [P’s truth on Monday], is factually all about
the italicized matter. This suggests that her rational credence at tm in R is 0.3 given (i) and
(iii).
What if B doesn’t know at tm that it is Monday? I think the natural expansion of
the above discussion is to add the condition that it is Monday. Hence, on the condition
43434343
that (i) a form of precipitation occurs now, (ii) it is Monday, and (iii) B’s past credence
function at ts was such that Cts(R on Monday /P on Monday)=0.3, the rational credence in
raining is 0.3 according to the idea of the Conditional Expert Principle. Formally,
Ctm(R/P&Vm&Ct
s(R in vm/P in vm)= 0.3)=0.3, where vm is Monday, and so Vm is that it is
Monday.
We can extract a general idea from this example. On the condition that (i)
observation proposition Ei is presently true, (ii) it is vj, and (iii) the previous credence was
such that Cn(X in vj/Ei in vj)=r where the truth-values of X and Ei are invariant within each
vj, the credence of X must be r. More formally,
(Shifted Backward Reflection) Cn+1(X/Ei&Vj&Cn(X in vj/Ei in vj)=r)=r for all x
such that Cn+1(Ei&Vj&Cn(X in vj/Ei in vj)=r)>0, where Ei&Vj is a member
of the time-observation partition at tn+1.
Of course, Shifted Rigidity follows from Shifted Backward Reflection if the agent
remembers her past credence distributions with perfect confidence.27
Hence, the idea of SR is best understood when we stipulate that from an agent’s
present point of view, the agent herself at the previous moment is an expert only lacking
the information confirmed by her present experience. For if that stipulation is true, it will
be rational for the agent to coordinate her present credence distribution with her previous
27 Let r be Cn(X in vj/Ei in vj). If the agent remembers her past credence function with perfect confidence,
Cn+1(Cn(X in vj/Ei in vj)=r)=1. Then, Cn+1(X/Ei&Vj)=Cn+1(X/Ei&Vj&Cn(X in vj/Ei in vj)=r)=(by Shifted
Backward Reflection) r=Cn(X in vj/Ei in vj). Done.
44444444
one according to SR. Since the stipulation seems to be true and SR entails SJC, we have
an argument for SJC.
F. Sleeping Beauty and Shifted Jeffrey Conditionalization
What answer does SJC give to the SB problem? In this section, I make three claims: First,
it is disprovable under SJC that SB’s credence in H stays the same from Sunday to
Monday. Second, it is provable under SJC that her credences in H1 and T1 are the same
on Monday. Third, it is not obviously provable or dis-provable that her credences in T1
and T2 are the same on Monday.
In my discussion in this section, the target tensed propositions will be H, T, H1,
T1, and T2, the evidence will be W, and the partition of intervals will be Monday,
Tuesday. The truth-values of these target tensed propositions and evidence are invariant
within Monday and within Tuesday. Hence, we can apply SJC to SB’s updating from
Sunday to Monday with the time-observation partition W&MON, W&TUE.28
Equipped with SJC, I criticize the Halfers’ thesis (1), which asserts that with no
relevant new evidence, no one can rationally change her credence in a genuine
proposition. I show that the SB problem is a counterexample of this thesis. Look at this
instance of SJC:
28 Why do we use SJC with exactly this fine-grained time-observation partition? On the one hand, the truth-
value of W is not invariant within Monday+Tuesday (0:00 AM on Monday to 11:59 PM on Tuesday).
Hence, you cannot use SJC with the time-observation partition W&(MON∨TUE). On the other hand, the
truth-value of W is invariant within each of Monday AM, Monday PM, Tuesday AM, Tuesday PM.
Hence, you can use SJC with the time-observation partition W&MONAM, W&MONPM, W&TUEAM,
W&TUEPM, but it will generate the same result as using SJC with W&MON, W&TUE. In sum, it
violates SJC’s proviso to use a more coarse-grained time-observation partition, and it is a waste of
calculating effort to use a more fine-grained time-observation partition.
45454545
(8) Cm(H) = Cs(H on Monday/W on Monday)*Cm(W&MON)+
Cs(H on Tuesday/W on Tuesday)*Cm(W&TUE).29
On the one hand, the first conditional credence is 1/2, which is equal to her credence on
Sunday in H. For “W on Monday” in the first conditional credence phrase is redundant
because she fully expected on Sunday night that she would wake up on Monday;
furthermore, “on Monday” in the resulting unconditional credence phrase also would be
redundant because H is a genuine proposition whose truth-value is insensitive to time. On
the other hand, the second conditional credence is 0. For waking up on Tuesday means
that the coin lands tails. In sum, her credence on Monday in H is the weighted average of
1/2 and 0, where the weights are her credences on Monday in W&MON and in W&TUE.
Since SB cannot rationally rule out either possibility, 0<Cm(H)<1/2. Because W doesn’t
seem to be new evidence relevant to H, this is a counterexample of (1).
In general, this explains how (1) can be violated by an SJC-er: Even when an
agent’s total evidence E is not new and relevant to X, she may change her credence in X.
For if she is uncertain whether her temporal location is in v1 or in v2, she is uncertain
between two time-observation propositions, E&V1 and E&V2. Even if the certain
evidence, E, is not new in that she fully expected that E would be true, one or both of
E&V1 and E&V2 can be new in that she didn’t fully expect E’s truth in v1 and/or in v2. In
such a case, an SJC-er can change her credence in X because arithmetically, it’s the
newness/oldness of the time-observation propositions, not of the evidence, that decides
29 In this section, I will use “… on Monday” as the abbreviation of “… is true (at every moment) on
Monday”; likewise for “… on Tuesday.”
46464646
whether it is rational for the agent to change her credence in X. (This point generalizes to
any n-case.)
Second, I defend the Thirders’ thesis (2): When SB wakes up on Monday, her
credence in H1 is the same as that in T1. As I discussed in Section A, Thirders have
defended this thesis with the intuition that SB’s credences in H and T would be equally
1/2 if she received additional information that it is Monday; hence, the actual conditional
credences in H and T, given MON, also should be 1/2. Arithmetically, this leads to thesis
(2): Cm(H1)=Cm(T1). I consider this to be a sound argument.
However, it would be more convincing if they could derive SB’s credence
distribution on Monday from that on Sunday, rather than from her credence distribution
in a counterfactual situation. With SJC, that derivation is possible. Here are its two
instances for SB’s credences at m in H1 and T:
(9) Cm(H1) = Cs(H1 on Monday/W on Monday)*Cm(W&MON)+
Cs(H1 on Tuesday/W on Tuesday)*Cm(W&TUE).
(10) Cm(T1) = Cs(T1 on Monday/W on Monday)*Cm(W&MON)+
Cs(T1 on Tuesday/W on Tuesday)*Cm(W&TUE).
According to (9), the credence on Monday in H1 is the weighted average of the
conditional credences on Sunday in [H1’s truth on Monday] given [W’s truth on Monday]
and in [H1’s truth on Tuesday] given [W’s truth on Tuesday]. The first conditional
credence is 1/2. For [H1’s truth on Monday] is equivalent to H’s truth simpliciter, and SB
fully expected W to be true on Monday; hence, it equals the credence on Sunday in H, 1/2.
47474747
The second conditional credence is 0. For H1 cannot be true on Tuesday. Hence, SB’s
conditional credence on Monday in H1 is the average of 1/2 and 0, the former weighted
by the credence in W&MON and the latter by that in W&TUE. As a result, Cm(H1)=1/2
Cm(W&MON). Similarly, it follows from (10) that Cm(T1)=1/2Cm(W&MON). Therefore,
Cm(H1)=Cm(T1).
We already established the truth of the Lesser view. However, thesis (2) provides
an alternative proof: Remember that H1, T1, and T2 exclusively exhaust all possibilities
open to SB at the moment of wakeup on Monday. She cannot rule out T2 and so
Cm(H1)+Cm(T1)<1. Since Cm(H1)=Cm(T1), Cm(H1)<1/2. Because H1 is the only possibility
in which the coin lands heads, Cm(H)<1/2.
Given this, what is the precise value of SB’s credence in H when she wakes up?
As discussed in Section A, the answer will be 1/3 if thesis (3) is true: When SB wakes up
on Monday, her credence in T1 is the same as that in T2. Unfortunately, there is no
obvious way to prove or disprove this thesis by SJC. (Try it.)
Elga says that we can prove (3) by a principle of indifference:
…even a highly restricted principle of indifference yields that you ought then to have equal
credence in each. (Elga 2000, 144.)
No doubt, “each” refers to each of T1 and T2. Thus, Elga is arguing here that (3) follows
from a highly restricted principle of indifference. In response to this, I ask two questions:
First, does (3) really follow from his principle of indifference? Second, is his principle of
indifference possibly true?
48484848
Both questions are hard to answer because Elga did not provide an explicit
formulation of his principle of indifference in his paper (2000). Fortunately, he provided
a formulation of that principle in a more recent paper (Elga 2004, 387): Define a centered
world to be a maximally consistent centered proposition.
(INDIFFERENCE) For any centered worlds X and Y, a rational agent ought to assign the
same credence to X and Y if (i) they are associated with the same possible world (i.e. for
some possible world W, X and Y both imply that W is the actual world) and (ii) they
represent epistemic situations that are subjectively indistinguishable (i.e. whichever of X
and Y is true of you, your experience will be exactly the same).
At first glance, this principle appears to entail (3): Assume that SB fully knows
everything about her world except whether the coin lands heads or tails. Under this
assumption, we can think as if T1 and T2 are centered worlds satisfying (i) and (ii). Hence,
it follows that Cm(T1)=Cm(T2). However, Weatherson (2005) criticizes this approach: First,
he says, we cannot infer (3) from INDIFFERENCE without the above assumption. For
define S to be the set of possible worlds such that for each W in S, the actuality of W is
compatible with T1 and T2, and consider the case where S is uncountably large. In this
case, it is perfectly coherent for SB to assign the same credence to any two centered
worlds associated with the same possible world in S but different credences to T1 and T2.
(For more details, see Weatherson (2005, pp. 615-616).)
Second, INDIFFERENCE is incompatible with Countable Additivity. For
assume a possible world W containing an infinite but countable number of agents who are
49494949
in the epistemic situations that are subjectively indistinguishable; if Countable Additivity
is true, it is incoherent to assign the same credence to the centered worlds representing
these agents’ epistemic situations in W. (For more details, see Weatherson (2005, pp.
619-621).) In sum, INDIFFERENCE entails (3), but it does so only under a highly
unlikely assumption, and INDIFFERENCE is incompatible with the widely accepted
axiom of Countable Additivity. Hence, the validity and factual correctness of Elga’s
argument are both questionable. That said, I leave it as an open question whether there
exists a consistent principle of indifference entailing (3).30
So far, I have disproved (1) and proved (2) by SJC. I have pointed out that (3) is
not obviously provable or disprovable under SJC. Consequently, I reject the Halfer view
and accept the Lesser view. However, I leave it as an open question whether the exact
value of SB’s credence in H is fixed by a principle of indifference or any other
consideration.
G. Conclusion
In this chapter, I have defended the Lesser view of the SB problem by SJC, a new
principle for updating de nunc credences. My discussion not only defends the Lesser
view, but it also provides a clue for what has gone wrong with the Halfer view. Read
Elga’s following comment on SB’s increasing credence in H:
30 However, I am skeptical. First, Elga points out that INDIFFERENCE leads to the weird result of “a brain
race,” although he bites the bullet (Elga 2004, p. 394). Second, Weatherson argues (convincingly in my
opinion) that INDIFFERENCE is not motivated because the intuition behind it is better captured in the
framework of imprecise credences (Weatherson 2005, p. 624). In my opinion, any principle similar to
INDIFFERENCE is likely to share these problems.
50505050
This belief change is unusual. It is not the result of your receiving new information. … So what
justifies it? The answer is that you have gone from a situation in which you count your own
temporal location as irrelevant to the truth of H, to one in which you count your own temporal
location as relevant to the truth of H. (Elga 2000, 146)
SJC provides a good explanation as to why Elga’s statement is correct: It follows from
SJC that Cm(H)=Cs(H/W on Monday)Cm(W&MON)+Cs(H/W on Tuesday)Cm(W&TUE)=
1/2Cm(MON)+0Cm(TUE)∈(0,1/2). This decrease was possible because, given that SB is
awake, TUE confirms [W’s truth on Tuesday], which is negatively relevant to the truth of
H. Since SB’s earlier knowledge that it was Sunday was not relevant to H in this way, she
has gone from [a situation in which she counts her temporal location as irrelevant] to [one
in which she counts it as relevant]. In this case, even if the evidence is old and irrelevant
to X, it is not sufficient for the rationality of not changing the credence in X.
Since the debate between Halfers and Thirders has been due not to the lack of
supporting arguments but to the failure of each side to point out the other side’s problem,
this is a good achievement. However, the even greater accomplishment seems to be the
updating principle itself. David Lewis wrote that the rule for updating de se credences is
formally identical to the rule for updating de dicto credences.31
This is wrong. As I have
demonstrated in this chapter, SJC is a good candidate for a rational principle for updating
the narrower category of de nunc credences. Obviously, SJC needs further
31 Lewis (1979) writes: “Then it is interesting to ask what happens to decision theory if we take all attitudes
as de se. Answer: very little. We replace the space of worlds by the space of centered worlds … All else is
just as before.” (Lewis 1979, pp. 533-4).
51515151
generalizations, but at least it initiates a good starting point from which we can find the
universal rule for updating de se credences.
52525252
CHAPTER III
UPDATING WITH A SEQUENCE OF OBSERVATIONS
A. Introduction
In Chapter II, I argued that a rational agent updates her credence in a tensed proposition
by SJC. In formulating SJC, I introduced this proviso: In updating from a past credence
function Cn at tn to the present credence function Cn+1 at tn+1, the agent observes nothing
between tn and tn+1.
Due to this proviso, SJC is not versatile. Compare it with SC, the traditional rule
for de dicto updating. According to the de dicto version of SC, if an agent observes E1, E
2,
E3, …, E
m during (tn,tn+m], she can simply conditionalize upon the conjunction E
1&E
2&
E3&…&E
m to update her credence in any proposition X (where X, E
1, E
2, E
3, …, E
m are
genuine propositions). This fact suggests that as a result of observing E1, E
2, E
3, …, E
m,
she learns E1&E
2&E
3&…&E
m.
Unfortunately, the same cannot be said for de se observations. Suppose that Jane
watches a cloudy sky first and a clear sky later, from 4PM to 5PM. What does she learn
as a result? If it is the conjunction of watching a cloudy sky and watching a clear sky,
then she must have learned the same thing as what she would have learned if she had
watched a clear sky first and a cloudy sky later during that period. Of course, this is
absurd. In general, when a given agent’s observations are de se, their order is important
in determining what the agent has learned as a result of those observations. Therefore,
what she learns cannot be just the conjunction of those observations.
53535353
For this reason, I chose to apply SJC only under the proviso that the agent
observes nothing during (tn,tn+1). When that proviso is satisfied, what she learns as the
result of her observations during (tn,tn+1] is simply what she observes at tn+1. In that way,
we did not have to worry about the order of observations.
However, this proviso is too restrictive. For example, think about this version of
the SB problem: SB problem 1. On Sunday, SB knows that she will experience the
following experiment. One minute later, she is put to sleep by a group of evil
experimenters. Then, they toss a fair coin. Case 1: (H) The coin lands on heads. In this
case, they awaken her only once on Monday. Case 2: (T) The coin lands on tails. In this
case, they awaken her twice, the first time on Monday and the second time on Tuesday;
between the two awakenings, they inject her with a drug that erases her memory of the
first awakening. In either case, one minute after she wakes up on Monday, she is told that
it is Monday. The experiment ends on Wednesday when she wakes up with the memory
of the previous awakening.
Let s be SB’s last conscious moment on Sunday, m be the moment of her wakeup
on Monday, and m+ be that of her being told that it is Monday. During (s,m+), she
observes W (“SB wakes up with the memory of Sunday as the last memory”). Since this
violates the proviso, SJC does not apply to SB’s credal transition from s to m+. One may
say that this is not a big problem because we can apply SJC to her credal transition from s
to m and then to her credal transition from m to m+. However, it is clear that it would be
better if we had an updating rule free from this kind of restriction.
In this chapter, I will suggest that if an agent makes a sequence of de nunc
observations E1, E
2, E
3, …, E
m during (tn,tn+m], she may update her de nunc credence in a
54545454
tensed proposition X by using a new updating rule that I call “Sequential Shifted Jeffrey
Conditionalization” (hereafter: SSJC). I will defend SSJC to some extent, but I will also
point out that SSJC does not apply to every case. Hence, I will try to provide some
criterion to distinguish the cases in which SSJC is a rational updating strategy from those
in which it is not.
B. Review of SC and SJC
In this section, I will review [the de se version of SC] and SJC. In particular, I will
explain how the latter solves a problem of the former.
Consider this case: Example 1. On Monday, Jane listens to the radio news,
which reports that today is Monday and that if there is any form of precipitation in
Boston on Tuesday, it will be rain. Consequently, Jane is sure to the degree of 0.9 that
today is Monday, but she does not completely trust the news, and so she assigns the
credence of 0.1 to the possibility that today is Tuesday. Conditional on its being Tuesday,
she ascribes no authority to the news. Hence, (i) she believes to the degree of 0.5 that [it
rains in Boston today] given that [Today is Tuesday and there is a form of precipitation in
Boston today].32
Conditional on its being Monday, Jane regards the weather news as
authoritative. Since she is pretty sure that it’s Monday, she ascribes some authority to the
news, even unconditionally. Thus, (ii) she believes to the degree of 0.8 that [it rains in
Boston on Tuesday] given that [there is a form of precipitation in Boston on Tuesday].
After listening to the news, she is put to sleep and stays in that state until next morning.
Waking up on Tuesday, Jane is told by her guru that today is Tuesday and that there is a
32 Let’s suppose that she knows that if any form of precipitation in Boston on Tuesday, it will be a rain or
snow, and she will be neutral between the two possibilities conditional on any form of precipitation there
on Tuesday.
55555555
form of precipitation today. At that moment, how confident will Jane be that it rains
today?
If Jane is a de se SC-er, she will be confident to the degree of 0.5 that it rains
today. If Jane is an SJC-er, she will be confident to the degree of 0.8 that it rains today.
Why is the former view wrong? To answer, let me formulate the de se version of SC
again: For simplicity, suppose that the agent B makes no observation during (tn,tn+1). Let
X be a centered proposition and E be the de se evidence that the agent B receives at tn+1.
Then,
(1) Cn+1(X)=Cn(X/E)
where Cn and Cn+1 are B’s credence functions at tn and tn+1. The goal from B’s point of
view at tn+1 is to find the rational credence that X is true now given the evidence that is
true now, where “now” refers to tn+1. However, the right-hand side of (1) denotes the
agent’s earlier credence that X was true then given that E was true then, where “then”
refers to tn. In this sense, if B is a de se SC-er, she comes to set her present credence in X
by consulting an outdated conditional credence. I will call this problem “the outdated
conditional credence problem.” Note: this was not a problem for the de dicto version of
SC because the target proposition and evidence have fixed truth-values in the de dicto
framework.
Consider Example 1 again. For brevity, let R be the tensed proposition that it
rains in Boston today, TUE be the tensed proposition that today is Tuesday, and P be the
tensed proposition that there is a form of precipitation in Boston today. TUE&P is Jane’s
56565656
total evidence on Tuesday. If she is a de se SC-er, her credence on Tuesday in R will be
her credence on Monday in R given TUE&P. From Jane’s point of view on Tuesday,
when she judges how likely it is that R is true now, she is consulting her credence in R’s
then truth given TUE&P’s then truth, where “now” refers to Tuesday and “then” refers to
Monday. Intuitively, this means that she comes to set her credence in R by consulting a
temporally mismatching conditional credence.
Fortunately, SJC helps us overcome the outdated conditional credence problem.
For simplicity, focus upon its sub-principle SSC: Suppose that the agent B receives no
evidence during (tn,tn+1) and knows that the time is v. Let X be a tensed proposition and E
be the de nunc evidence that the agent B receives at tn+1 (where the truth-values of X and
E are invariant within v). Then,
(2) Cn+1(X)=Cn(X in v/E in v)
where Cn and Cn+1 are B’s credence functions at tn and tn+1. In words, B sets her present
credence in X to be her previous credence that [X is true in v] given that [E is true in v].
This is intuitively reasonable, because from B’s point of view at tn+1, X is true iff “X is
true in v” is true at whatever time, and E is true iff “E is true in v” is true at whatever time.
(More about this point below.)
In Example 1, if Jane is an SSC-er, her credence on Tuesday in R will be 0.8.
For her credence on Tuesday in R will be her credence on Monday in [R’s truth on
Tuesday] given [TUE&P’s truth on Tuesday]. From her point of view on Tuesday,
57575757
(3) R is true now iff [R’s truth on Tuesday] holds at whatever time, and
(4) TUE&P is true now iff [TUE&P’s truth on Tuesday] holds at whatever
time.
These facts have two consequences: First, given these equivalences, it seems okay on
Tuesday for Jane to evaluate the credal impact of TUE&P on R by evaluating that of
[TUE&P’s truth on Tuesday] on [R’s truth on Tuesday].33
Second, there is no problem of
temporal mismatch in setting her credence on Tuesday in R with evidence TUE&P by
consulting her previous credence in [R’s truth on Tuesday] given [TUE&P’s truth on
Tuesday]. For the bracketed propositions are genuine propositions, which have fixed
truth-values. Therefore, this application of SSC is free from the outdated conditional
credence problem.
To generalize the above discussion, I introduce the following definition: Let X be
any tensed proposition and v be a temporal interval. Additionally, let V be the tensed
proposition that it is v now. Then,
(5) The de-indexicalization of X under V is the genuine proposition that X is
true throughout v.
Given (5), we can verbalize (2) into this claim:
33 By “credal impact,” I mean the quantity of the force of evidence or observation that increases or
decreases the agent’s credences. While this definition is not a good example of philosophical clarity, I am
not alone in using this somewhat opaque notion. See Lewis (1980; 272) for an example.
58585858
(6) B’s credence at tn+1 in X=B’s credence at tn in [the de-indexicalization of X
under V] given [the de-indexicalization of E under V].
The point of de-indexicalization here is to convert the target tensed proposition X and de
nunc evidence E into the equivalent genuine propositions X′ and E′ so that B can judge
how probable X is given evidence E by checking her previous credence in X′ given E′. In
this process, de-indexicalization allows the agent to use her previous conditional credence
to set her present credence without the outdated conditional credence problem.
As we have seen, SSC solves the outdated conditional credence problem by de-
indexicalizing the target tensed proposition and evidence. We can regard SJC as a
generalization of SSC where the agent is unsure about what observation she has made
and what time it is. The next question is, “How can we generalize the idea of de-
indexicalization for when the agent receives a sequence of evidence?”
C. Strategy
To answer the above question, I want to discuss SB’s credal transition from s to m+ as an
example. I begin by asking three questions: First, what does SB learn during (s,m+]?
Second, is there a genuine proposition equivalent to what she learns as a result of her
observations during that interval? Third, what is the correct way to update her credence in
H from s to m+ by making use of what she learns?
In answer to the first question, SB observes W at m and observes MON at m+. As
a result, she comes to fully believe at m+ that (E) W was previously true and MON is
presently true. In sum, E is what she learns as a result of her observations during (s,m+].
59595959
In answer to the second question, SB knows at m+ that it was previously Monday
and that the truth-value of W is invariant within Monday. Thus, she will accept this bi-
conditional at that moment:
(7) “W was previously true” is true now iff “W is true on Monday” is true at
whatever time.
Because she also knows at m+ that it is presently Monday and the truth-value of MON is
invariant within Monday, she will accept
(8) MON is true now iff “MON is true on Monday” is true at whatever time,
at m+. Now, let D be the genuine proposition that W is true on Monday and MON is true
on Monday. From the point of view at m+, D and E are equivalent because their
conjuncts are equivalent. Since D is a genuine proposition, there exists a genuine
proposition that is equivalent at m+ to E.
And now to the third question: what is the correct method of SB’s updating her
credence in H from s to m+? Since E is equivalent to D, she can evaluate the evidential
impact of E upon H by evaluating that of D on H. Hence,
(9) Cm+(H)=Cs(H/D)=Cs(H/MON is true on Monday&W is true on Monday).34
34 Since SB is sure of E at m+, Cm+(H)= Cm+(H/E). Thus, if E’s impact at m+ on H can be measured by
checking D’s impact at s on H, then (10) is derived.
60606060
It is trivial that MON is true on Monday, and on Sunday SB fully expected to wake up on
Monday. Hence,
(10) Cm+(H)=Cs(H)=1/2.
This result complies with the traditional Thirder view (Elga 2000).
Let us suppose that (9) is the correct way for SB’s updating from s to m+. If we
can generalize it, perhaps we can find a model for an agent’s updating from tn to tn+m
where the agent makes observations many times during (tn,tn+m]. The remaining question
is “How?”
In the rest of this chapter, I will proceed in the following order: In Section D, I
will present two new updating principles. In Section E, I will construct an argument for
these new principles by utilizing the idea of the Conditional Expert Principle, as I did in
Chapter II. In Section F, however, I will argue that those updating principles lead to
mutually inconsistent results when applied to the SB problem. In Section G, I will
provide a diagnosis for the problem. In Section H, I will formulate weaker versions of the
two principles suggested in Section D.
D. Updating with a Sequence of Observations
In this section, I will discuss how to update de nunc credences after making a sequence of
observations. First, I will expand our formal language L into a larger language L′, and
then I will introduce a few new definitions. Next, I will present two de nunc updating
61616161
principles by using the new language and definitions. Finally, I will illustrate how these
principles work in several examples.
I begin by expanding language L, which I constructed in the last chapter, into a
new language L′:
First, every expression in L is also a legitimate expression in L′.
Second, L′ includes an indexical “prev.” Remember the assumption that I made
in the previous chapter (see footnote 14): The agent makes observations only at a
countable number of times, say, ... tn-2, tn-1, tn, tn+1, tn+2, tn+3, .... Let us call those moments
“epistemic moments.” From now on, I will use the numeric subscripts to indicate the
order and contiguity of the epistemic moments. In other words, for any n,m∈N, tn is an
earlier epistemic moment than tm iff n<m and tn is the last moment the agent observes
anything before tn+1. Then, here is the meaning schema for “prev”:
(11) At tn, “prev” refers to tn-1.
English has no precise counterpart to “prev” in L′, but I will often use “the previous
moment” to mean the same thing.
Third, L′ includes an indexical “pres.” Here is the meaning schema:
(12) At tn, “pres” refers to tn.
Hence, “pres” in L′ amounts to “the present moment” in English.
62626262
Fourth, L′ includes (somewhat artificial) indexicals “prevk,” where k≥0. Here is
the meaning schema for “prevk”:
(13) At tn, “prevk” refers to tn-k.
As a result, “prev0” refers to the same epistemic moment as “pres” does, and “prev1”
refers to the same moment as “prev” does. This finishes our expansion of L into L′.
Next, I define two important notions. In order to do so, I ask the following
questions: First, if an agent makes a sequence of observations, what does she come to
learn at the end? Second, assuming an answer to the previous question, is there a genuine
proposition equivalent to what she learns?
Focus on the first question. Let me precisify it first: Suppose that an agent B
observes E1, E
2, E
3, ... E
m, and nothing else during interval (tn,tn+m]. Then, what does B
come to learn at tn+m as the result? Consider this answer:
(14) E1&E
2&E
3&...&E
m.
Unfortunately, this suggestion is inadequate when E1, E
2, E
3,... E
m are irreducibly tensed
propositions. For (14) ignores the temporal gaps among E1, E
2, E
3,... E
m. To understand
this point, think about the following example: Example 2. Let E1 be the tensed
proposition that the sky is cloudy now and E2 be the tensed proposition that the sky is
clear now. Suppose that Jane observes E1 at 4:30 PM and E
2 at 5:00 PM during (4:00 PM,
5:00 PM], and nothing else during that period. Then, there is a serious problem if we
63636363
regard E1&E
2 as what Jane learns as the result. For, at whatever time it is evaluated,
E1&E
2 entails that the sky is clear and cloudy at the same time. Since Jane’s perfectly
normal experience cannot lead to such absurdity, E1&E
2 is not what she comes to learn as
the result of her observations during (4:00 PM, 5:00 PM].
Instead, I suggest that what an agent B learns as the result of her observations
E1,E
2,E
3,..., E
m during (tn,tn+m] is
(15) (E1 at prevm-1)&(E
2 at prevm-2)&(E
3 at prevm-3)&...&(E
m at prev0)
in L′. Why? Note that (15) is the translation to L′ of the following expression in (my
dialect of) English:
(16) E1 was true m-1 epistemic moments ago&
E2 was true m-2 epistemic moments ago&
...
Em
is true 0 epistemic moment ago (or now).
Remember that epistemic moments are the moments at which the given agent observes
anything. If B has perfect memory as is usually assumed, then B will remember that she
observed E1 m-1 epistemic moments ago, E
2 m-2 epistemic moments ago, etc. In general,
B will remember that she observed Ek m-k epistemic moments ago, for any k such that
1≤k≤m. Thus, as a result of having observed E1, E
2, E
3,... E
m, she comes to learn the
tensed proposition expressed by (15) or (16). For example, Jane will believe at 5:00 PM
64646464
that (E1) “the sky is cloudy” was true when she observed anything last time, and (E
2) “the
sky is clear” is true now. This belief is expressed in L′ by “(E1 at prev1)&(E
2 at prev0)” or
“(E1 at prev)&(E
2 at pres).” In my opinion, this is what Jane learns as the direct result of
her observations during (4:00 PM, 5:00PM].
In general, I suggest the following definition:
(17) E is the agent B’s sequential total observation during (tn,tn+m] iff
(i) B observes E1, E
2,... E
m and nothing else during (tn,tn+m], and
(ii) E=(E1 at prevm-1)&(E
2 at prevm-2)&...&(E
m at prev0).
(Be careful: While “E” carries the increasing subscripts from 1 to m, “prev” carries the
decreasing subscripts from m-1 to 0.)
Now, let’s focus on the second question. Again, I first precisify the given
question: If (E1 at prevm-1)&(E
2 at prevm-2)&...&(E
m at prev0) is B’s sequential total
observation during (tn,tn+m], is there a genuine proposition equivalent to that
observation?35
To answer this question, first, I expand the notion of de-indexicalization defined
in Section B: Let E be a tensed proposition and v be a temporal interval such that the
truth-value of E is invariant within v. Thus, V is the tensed proposition that it is v now.
Then,
35 When it results in no confusion, I shall mix target language, L′, with the meta-language, English. The
same applies to the rest of this dissertation.
65656565
(18) The de-indexicalization of [E at prevk] under [V at prevk ] is [E in v].
Here is the core idea: Under the hypothesis that it was v k epistemic moments ago, E was
true k epistemic moments ago iff E was true in v.36
Hence, under the temporal hypothesis
[V at prevk], the agent will judge that [E at prevk] is true iff the genuine proposition [E in
v] is true at whatever time.
Next, I expand the notion of de-indexicalization for the case where an agent
makes a sequence of observations during an interval:
(19) The sequential de-indexicalization of
(i) (E1 at prevm-1)&(E
2 at prevm-2)&...&(E
m at prev0)
is
(ii) (E1 in v
1)&(E
2 in v
2)&...&(E
m in v
m)
under the temporal hypothesis
(iii) (V1 at prevm-1)&(V
2 at prevm-2)&...&(V
m at prev0)
where the truth-value of Ek is invariant within v
k for any k∈1,...,m.
This looks complicated but the core idea is the same as before: Under the temporal
hypothesis (iii), (i) is true now iff (ii) is true at whatever time. Given this conditional
equivalence, the agent can evaluate the credal impact of (i) by evaluating that of (ii) when
she knows that (iii) is true. For an agent’s sequential total observation is always
36 Suppose [V at prevk]. Clearly, this supposition entails that prevk ∈v. By the definition of “in,” [E in v]
implies [E at prevk]. Since we are assuming that the truth-value of E is invariant within v, [E at prevk]
implies [E in v]. Done.
66666666
equivalent to its sequential de-indexicalization under the correct temporal hypothesis
(about the relevant epistemic moments).
Having defined these notions, I am ready to formulate my first updating rule in
this chapter, “Sequential Shifted Strict Conditionalization”: Consider a sequence of
observations E1, E
2, ...E
m and a sequence of intervals v
1, v
2, ... v
m. Assume that the truth-
value of X is invariant within vm
and that of Ek is invariant within v
k for each k∈1,...,m.
Then, for any tensed proposition X,
(SSSC) Cn+m(X)=Cn(X in vm
/(E1 in v
1)&(E
2 in v
2)&...&(E
m in v
m))
if B is sure at tn+m that for each k∈1,...,m, [Ek was/is true and it was/is
vk] at the m-k epistemic moments earlier time,
where Cn and Cn+m are B’s credence functions at tn and tn+m. Less formally, we can
rewrite SSSC in this way: Let E be (E1 at prevm-1)&...&(E
m at prev0) and V be (V
1 at
prevm-1) &...&(Vm
at prev0). Then,
(SSSC) Cn+m(X)=Cn(the de-indexicalization of X under Vm
/the sequential de-
indexicalization of E under V), if B has certainly learned until tn+m that E&V is true.
In my opinion, this is a reasonable rule for de nunc updating. It is free from the outdated
conditional credence problem, as the target tensed proposition and sequential total
observation are converted to the genuine propositions whose truth-values are fixed.
67676767
To see how SSSC works, consider this example: Example 3. Let F be the tensed
proposition that sparrows are flying away from Washington, DC today, let R be that
animals are running away from Washington, DC today, and let Q be that there is an
earthquake in Washington, DC tomorrow. On Sunday, Jane’s credence in [Q’s truth on
Tuesday] given [F’s truth on Monday]&[R’s truth on Tuesday] is 0.3. On Tuesday, she
remembers that she observed sparrows flying away from Washington, DC and that it was
Monday, and she is sure that she is observing animals running away from Washington,
DC and it is Tuesday. Jane did not have any other relevant evidence from Monday to
Tuesday. In this case, to what degree should Jane believe in Q on Tuesday?
Intuitively, her credence on Tuesday in Q must be 0.3. For she believed that [Q
would be true on Tuesday] to the degree of 0.3 conditional on the assumption that [F
would be true on Monday] and [R would be true on Tuesday], and her observations on
Monday and Tuesday exactly confirm this assumption. SSSC captures this intuition, as it
is an instance of SSSC for this example that CTUE (Q)=CSUN (Q in vT/(F in vM)&(R in
vT))=0.3, where vM is Monday and vT is Tuesday and CSUN and CTUE are Jane’s credence
functions on Sunday and Tuesday.
Next, let’s think about how to generalize SSSC for the following cases: The
agent updates from her old credence function at tn to a new credence function at tn+m, but
she is unsure what observations she has made and/or what times it has been after the m
epistemic moments earlier time (which we know to be tn). In this case, SSSC does not
apply because its proviso is not satisfied. So, what is the rational way for the agent to
update her credence?
68686868
Although SSSC does not apply to such a case, it provides an important clue for
the answer: Let Eo&Voo∈O be a partition such that Eo=(E1
o at prevm-1)&...& (Em
o at
prev0) and Vo=(V1
o at prevm-1)&...&(Vm
o at prev0) for each o∈O. I consider each Eo&Vo to
represent a possible scenario of what observations an agent B has made at what moments
(hereafter: a possible observational scenario). For simplicity, let O be 1, 2, ...p. By
SSSC:
Cn+m(X) would be
Cn(X in vm
1/D1) if B were sure at tn+m that E1&V1 is true,
Cn(X in vm
2/D2) if B were sure at tn+m that E2&V2 is true,
...
Cn(X in vm
p/Dp) if B were sure at tn+m that Ep&Vp is true,
where Do=(E1
o in v1
o)&...&(E1
o in v1
o) for each o∈O i.e., each Do is the sequential de-
indexicalization of Eo under Vo. Given these facts, it is natural that B’s credence at tn+m in
X is the weighted average of values on the right-hand sides of the above equations with
the weights coming from B’s credences at tn+m in Eo&Vo.
To formalize this idea, we need to have some preliminary jobs done. Consider a
partition &1≤k≤m((Eko&V
ko) at prevm-k)o∈O such that (i) Cn+m((E
ko&V
ko) at prevm-k)>0 for
each o∈O and (ii) ∑o∈OCn+m((Eko&V
ko) at prevm-k)=1 where Cn+m is an agent B’s credence
function at tn+m. (The intended interpretation of this partition is that each member is a
candidate for the observational scenario that B goes through during (tn+m,tn+m].) I will call
any member of this partition “(B’s) sequential time-observation proposition from tn to
69696969
tn+m.” If &1≤k≤m((Eko&V
ko) at prevm-k)o∈O also satisfies the condition that for each o∈O,
Cn(&1≤k≤m(Eko in v
ko))>0, then I will call the partition “(B’s) sequential time-observation
partition from tn to tn+m.”
Now, I am ready to formulate my next updating principle, “Sequential Shifted
Jeffrey Conditionalization”: Let &1≤k≤m((Eko&V
ko) at prevm-k)o∈O be an agent B’s
sequential time observation partition from tn to tn+m. Let Cn and Cn+m be B’s credence
functions at tn and tn+m. Assume that (#) the truth-value of X is invariant within vm
o and
that of Eko is invariant within v
ko, for any k∈1,...,m and o∈O. Then,
(SSJC) Cn+m(X)=
Σo∈OCn(X in vm
o/&1≤k≤m(Eko in v
ko))Cn+m(&1≤k≤m((E
ko&V
ko) at prevm-k)),
where Cn and Cn+m are B’s credence functions at tn and tn+m. Less formally: For each o∈O,
Eo=&1≤k≤m(Eko at prevm-k) and Vo=&1≤k≤m(V
ko at prevm-k). Clearly, Eo&Voo∈O=
&1≤k≤m((Eko&V
ko) at prevm-k)o∈O. Assume that (#) is satisfied. Then,
(SSJC) Cn+m(X)=the weighted average of Cn(the de-indexicalization of X under
Vm
o/the sequential de-indexicalization Eo under Vo) with the weights
coming from Cn+m(Eo&Vo),
70707070
where Cn and Cn+m are B’s credence functions at tn and tn+m. I believe this is a natural
generalization of SSSC for when the agent is not sure of what sequence of observations
she has made at what times.
To understand how SSJC works, think about this example: Example 4. Let F be
the tensed proposition that sparrows are flying away from Washington, DC today, RFROM
be that animals are running from Washington, DC today, RTO be that animals are running
to Washington, DC today, and Q be that there is an earthquake in Washington, DC
tomorrow. On Sunday, Jane’s credence in [Q’s truth on Tuesday] given [F’s truth on
Monday] & [RFROM’s truth on Tuesday] was 0.8, and her credence in [Q’s truth on
Tuesday] given [F’s truth on Monday]&[RTO’s truth on Tuesday] was 0.4. On Monday,
she observes that sparrows are flying away from Washington, DC and is certain of what
she is observing. On Tuesday, she observes that animals are running but is uncertain
whether they are running from Washington, DC or to Washington, DC. She observes
nothing else during the two days. On either day, she knows what day it is.
As a result, she is certain on Tuesday that she previously observed F and it was
Monday, and she is also certain on Tuesday that it is presently Tuesday. However, she is
uncertain about whether she is observing RAWAY or RBACK. Indeed, her credence on
Tuesday in [F’s previous truth and RFROM’s present truth]=0.5=her credence on Tuesday
in [F’s previous truth and RTO’s present truth]. I suggest that her credence on Tuesday in
Q will be 0.6. Look at Figure 6:
71717171
Figure 6: Evidential Uncertainty in Sequential Updating. The two rows represent two
possibilities about what Jane has observed on what days, during the last two days.
In this figure, the two rows dubbed “the first scenario” and “the second scenario”
represent the observational scenarios Jane might have gone through from Monday to
Tuesday. (i) On Sunday, Jane’s credence in Q’s truth on Tuesday was 0.8 given that [F
would be true on Monday and RFROM would be true on Tuesday]. On Tuesday, if
F&MON was previously true and RFROM&TUE is presently true, it will confirm the
bracketed condition. Therefore, Jane’s rational credence on Tuesday in Q is 0.8
conditional on F&MON’s previous truth and RFROM’s present truth. (ii) Similarly, Jane’s
credence on Tuesday in Q will be 0.4 conditional on F&MON’s previous truth and RTO’s
present truth. (iii) Therefore, her rational credence on Tuesday in Q is 0.6, the average of
0.8 and 0.4. I find this line of reasoning to be intuitive.
Interestingly, SSJC supports this intuitive claim. For it is an instance of SSJC
that CTUE(Q)=CSUN(Q in vT/(F in vM)&(RFROM in vT))*CTUE((F&VM at prev1)& (RFROM&VT
at pres))+CSUN(Q in vT/(F in vM)&(RTO in vT))*CTUE((F&VM at prev1)& (RTO&VT at pres))
=0.6, where vM is Monday and vT is Tuesday and CSUN and CTUE are Jane’s credence
functions on Sunday and Tuesday.
72727272
So far, so good. In the above example, the agent is uncertain of what she has
observed. But what if she is uncertain about what times it has been when she made those
observations? Consider this case: Example 5. Keep the meanings of “F,” “RFROM,” and
“Q” the same. On Sunday, Jane’s credence in [Q’s truth on Tuesday] given [F’s truth on
Monday]&[RTO’s truth on Tuesday] is 0.8, and her credence in [Q’s truth on Wednesday]
given [F’s truth on Tuesday]& [RTO’s truth on Wednesday] is 0.4. On that night, she
takes sleeping pills, but she realizes that she might have overdosed. If she did, she will
wake up on Tuesday. (Indeed, she didn’t overdose and will wake up and observe F on
Monday and RTO on Tuesday.)
On Monday, she wakes up and observes F. Then, she is immediately put to sleep
again, without taking any sleeping pill. (So she expects to wake up on the next day; for
the same reason, when she wakes up on the next day, she knows that only one day has
passed.) Waking up on Tuesday, Jane certainly knows that she previously observed F and
is presently observing RTO. However, since she is not sure that she didn’t overdose, she is
not sure that [it was previously Monday and it is presently Tuesday]; as far as she knows,
[it might have been Tuesday at the previous moment and it might be Wednesday now].
Consequently, her credence on Tuesday that [it was previously Monday & it’s Tuesday
now]=0.5=her credence on Tuesday that [it was previously Tuesday & it’s Wednesday
now].
73737373
Figure 7: Temporal Uncertainty in Sequential Updating. The two scenarios represent two
possibilities about what observations Jane has made on what days, for the last two days.
I claim that Jane’s credence on Tuesday in Q is 0.6. Why? Look at Figure 7:
In this figure, the two pairs of rows dubbed “the first scenario” and “the second scenario”
represent observational scenarios Jane might have gone through. (i) On Sunday, Jane’s
credence in Q’s truth on Tuesday was 0.8 given that [F would be true on Monday and
RTO would be true on Tuesday]. On Tuesday, if F&MON was previously true and
RTO&TUE is presently true, it will confirm the bracketed condition. Intuitively, her
rational credence on Tuesday in Q is 0.8 given that F&MON was previously true and
RTO&TUE is presently true. (ii) Similarly, Jane’s rational credence on Tuesday in Q is 0.4
given that F&TUE was previously true and RTO&WED is presently true. (iii) Therefore,
her credence on Tuesday in Q is 0.6, the average of 0.8 and 0.4. I find this to be intuitive
reasoning.
Importantly, SSJC supports this intuitive claim, as it is an instance of SSJC that
CTUE(E)=CSUN(E in vT/(F in vM)&(RTO in vT))*CTUE((F&VM at prev1)&(RTO&VT at
pres))+CSUN(E in vW/(F in vT)&(RTO in vW))*CTUE((F&VT at prev1)&(RTO&VW at pres))=
0.6, where vM, vT, and vW are Monday, Tuesday, and Wednesday, and CSUN and CTUE are
Jane’s credence functions on Sunday and Tuesday.
74747474
Finally, I want to provide more succinct formulations of SSSC and SSJC. For,
while the earlier formulations were good for the purpose of explanation, they often will
be too bulky for other purposes. So: Let Eo&Voo∈O be B’s sequential time-observation
partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k) and Vo=&1≤k≤m(V
ko at prevm-
k).37
Given this partition, we can define these notions: (i) Eo&Voo∈O is logically self-
optimal iff for each o∈O and k∈1,...,m, the truth-value of Eko is invariant within v
ko, (ii)
Voo∈O is logically optimal for X iff for each o∈O, the truth-value of X is invariant
within vm
o, and Eo&Voo∈O is logically optimal for X iff Eo&Voo∈O is self-optimal and
Voo∈O is optimal for X. Then,
(SSJC) Cn+m(X)=Σo∈OCn(X in vm
o/Do)Cn+m(Eo&Vo) if Eo&Voo∈O is logically
optimal for X,
where Do is the sequential de-indexicalization of Eo under Vo for each o∈O. Next, let
E&V be B’s sequential time-observation partition from tn to tn+m, where E=&1≤k≤m(Ek
at prevm-k) and V=&1≤k≤m(Vk at prevm-k). Then,
(SSSC) Cn+m(X)=Cn(X in vm/D) if E&V is logically optimal for X,
37 Note that Eo&Voo∈O=&1≤k≤m(E
ko&V
ko at prevm-k)o∈O. By the definition of sequential time-observation
partition, Eo&Voo∈O is such that (i) Eo=&1≤k≤m(Eko at prevm-k) and Vo=&1≤k≤m(V
ko at prevm-k), (ii) for each
o∈O, E1
o,…, Em
o are the candidates for B’s observations at t1,...,tn+m, (iii) for each o∈O, v1
o,…, vm
o are the
candidates for B’s temporal locations at t1,...,tn+m, (iv) Cn+m(Eo&Vo)>0 for each o∈O and
∑o∈OCn+m(Eo&Vo)=1, and (v) for each o∈O, Cn+m(Do)>0, where Do=&1≤k≤m(Eko in v
ko).
75757575
where D is the sequential de-indexicalization of E under V. Clearly, these reformulations
are equivalent to the original.
In this section, I discussed two sequential de nunc updating principles, SSSC and
SSJC. Just as SSC and SJC subsume the probabilistic inferential patterns we find to be
intuitive, SSSC and SSJC also subsume intuitive probabilistic reasoning patterns. In the
next section, I will defend these new principles by a modified version of Gaifman’s
Expert Principle.
E. A Defense of SSJC
In this section, I defend an intuitive principle that I call SSR. Since SSR entails SSJC and
SSSC, this will show that SSSC and SSJC are true of the cases similar to the example
given in this section.
First, consider this principle, which I call “Shifted Sequential Rigidity”: Let
Eo&Voo∈O be B’s sequential time-observation partition from tn to tn+m, where
Eo=&1≤k≤m(Eko at prevm-k) and Vo=&1≤k≤m(V
ko at prevm-k). Suppose that Eo&Voo∈O is
logically optimal for X. For each o∈O,
(SSR) Cn+m(X/Eo&Vo)=Cn(X in vm
o/Do),
where Do is the sequential de-indexicalization of Eo under Vo for each o∈O. Here, Eo&Vo
codifies what observations B has made in what times, and B knows that Do is equivalent
76767676
to Eo if Vo is true. It is not difficult to see that SSR entails SSJC.38
Hence, it suffices to
defend SSR.
But how do we defend it? Think about this principle, which I will call
“Sequential Shifted Backward Reflection”: Again, let Eo&Voo∈O be B’s sequential
time-observation partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k) and
Vo=&1≤k≤m(Vko at prevm-k). Then,
(SSBR) Cn+m(X/Eo&Vo&Cn(X in vm
o/Do)=r)=r if defined,
where Do is the sequential de-indexicalization of X under Vm
o. Approximately, SSBR is
the claim that it is rational for B to set her present credence in X to be the same as her
earlier credence in the de-indexicalization of X conditional on the sequential de-
indexicalization of Eo. Remember that it is standard to presuppose that the given agent
remembers her past credence functions with perfect confidence and correctness. Under
this presupposition, SSBR entails SSR.39
In sum, SSBR entails SSR, and SSR entails SSJC. Hence, it suffices to defend
SSBR. To do so, I appeal to a more general principle, which I call “the Tensed
38 Let Eo&Voo∈O be B’s sequential time-observation partition from tn to tn+m, where Eo=&1≤k≤m(E
ko at
prevm-k) and Vo=&1≤k≤m(Vko at prevm-k). Suppose that SSR is true, i.e., Cn+m(X/Eo&Vo)=Cn(X in v
mo/Do) for
each o∈O, where Do=&1≤k≤m(Eko at v
mo). Then, Cn+m(X)=Σo∈OCn+m(X&Eo&Vo)=Σo∈OCn+m(X/Eo&Vo)
Cn+m(Eo&Vo)= (by supposition) Σo∈OCn(X in vm
o/Do)Cn+m(Eo&Vo). Done.
39
Suppose that SSBR is true, i.e., Cn+m(X/Eo&Vo&Cn(X in vm
o/Do)=r)=r for any r∈[0,1] and o∈O. Let
r=Cn(X in vj/Do). By presupposition, Cn+m(Cn(X in vj/Do)=r)=1. Thus,
Cn+m(X/Eo&Vo)=Cn+m(X/Eo&Vo&Cn(X in vm
o/Do)=r)= (by supposition) r=Cn+m(Cn(X in vj/Do)=r). Done.
77777777
Conditional Expert Principle”: For any tensed propositions X and E, any genuine
propositions X′ and E′, and any temporal hypothesis V,
(TCE) C(X/E&V&prt(X′/E′)=r)=r if defined and these conditions are met:
(a) C is an agent B’s present credence function and prt is an agent Ex’s
credence function at t on which B depends to decide her present
credal opinion,
(b) Ex did not know at t whether E′ was true, B can presently access all
information that Ex had at t, and B has observed no other
information possibly except E.
(c) B presently knows that if V is true, [X is presently true iff X′ is true
at t] and [E is presently true iff E′ is true at t].
Suppose (a)-(c). By (a), it seems rational for B to have her credence in X somehow
coordinated with Ex’s credence distribution. But what is the rational way to do so?
To answer, think about the facts that follow from the following conditions
(under (a)-(c)): (i) E is presently true, (ii) V is presently true, and (iii) Ex’s credence at t in
X′ given E′ is r. It follows that E′ is true, which Ex did not know at t. Given this
informational impoverishment, B cannot rationally depend upon Ex’s unconditional
credences at t to set her present credences, conditional on (i)-(iii). Nevertheless, it is still
plausible that it is rational for B to use Ex’s credences conditioned upon E′. For whether
E is true is the only information that B presently has but that Ex might have lacked at t,
and E’s present truth is equivalent to E′’s truth at t. Hence, it will be rational for B to use
78787878
(iii) to set her present credence. Given this reasoning, it is rational that B’s present
credence in X is r on the condition that (i)-(iii) are true.
I claim that SSBR is a sub-principle of TCE. Consider the following instance of
SSBR, which involves B’s credal transition from tn to tn+m: Let Eo= &1≤k≤m(Eko at prevm-k),
Vo=&1≤k≤m(Vko at prevm-k), and Do=&1≤k≤m(E
ko in v
ko ). So Do is the sequential de-
indexicalization of Eo under Vo. Then,
(20) Cn+m(X/Eo&Vo&Cn(X in vm
o/Do)=r)=r if defined.
First, it is inevitable for B to set her present credence by depending upon her earlier
credence distribution. Second, Eo, if true, represents B’s total observation during (tn, tn+m],
which B could not access at tn. Third, if Vo (⊃Vm
o ) is true, X is true iff [X in vm
o] is true,
and Eo is true iff Do is true. Hence, B satisfies the provisos of TCE with respect to her
credence distribution at tn+m. Therefore, SSBR follows from TCE as a special case.
Consider this example: Example 6. Keep the meanings of “F,” “RFROM,” and
“Q” the same as in the earlier examples. At 9:00 AM on Tuesday (hereafter: t), Jane
regards herself at 9:00 PM on Sunday (hereafter: s) as an expert about local natural
phenomena. Then, what should her credence be at t in Q, given that (i) (E) F was
previously true and RFROM is presently true, (ii) (V) it was Monday previously and it is
Tuesday now, and (iii) her credence at s in Q’s truth on Tuesday was 0.7 given that F
would be true on Monday and RFROM would be true on Tuesday? My answer: it must be
0.7. Formally, Ct(Q/E&V&Cs(Q on Tuesday/D) =0.7)= 0.7, where D is the de-
79797979
indexicalization of E under V, i.e. the genuine proposition that F is true on Monday and
RFROM is true on Tuesday.
Why? Consider Figure 8:
Figure 8: Flying Birds, Running Animals, and Earthquake. The belief at the tail of an
arrow is true on the day of the belief’s column iff the event at the head of the arrow
occurs on the day of the event’s column.
First, to set her credence at t in Q, Jane seems to have no other choice but to use her
earlier credence distributions, possibly that at s. However, second, she must not simply
adopt her unconditional credence at s in Q as her credence at t in Q. This is because Jane
has observed some facts relevant to Q, such as birds flying away and animals running
away, after s. So what is the rational way for Jane to set her credence at t in Q by utilizing
her credence distribution at s? Third, here is a suggestion: Let E be (F at prev)&(R at pres)
and V be (MON at prev)&(TUE at pres). Then, Jane’s credence at t in Q is r on the
assumption that (i) E is presently true, (ii) V is presently true, and (iii) her credence at s in
80808080
[Q on Tuesday] is r given D, where D=(F on Monday)&(RFROM on Tuesday). For
although Jane cannot trust her credal judgments on Sunday for the reason mentioned, she
can still trust those conditioned upon D, which codifies her observations after Sunday.
Plus, Jane knows on Tuesday that if V is true, then Q is equivalent to [Q on Tuesday],
and so she can rationally set her credence in Q by checking her credence on Sunday in [Q
on Tuesday] conditioned upon D. (We need to substitute “must” for “can” in the last
sentence in light of the outdated conditional credence problem, which we discussed in
Section B.)
We can generalize this reasoning: Cn+m(X/Eo&Vo&Cn(X in vm
o/Do)=r)=r where
Eo=&1≤k≤m(Eko at prevm-k) is possibly the agent B’s sequential total observation during
(tn,tn+m], Vo=&1≤k≤m(Vko at prevm-k) is a temporal hypothesis about when B has made
observations during (tn,tn+m], and Do=&1≤k≤m(Eko in v
ko) is the sequential de-
indexicalization of Eo under Vo. It is easy to see that at tn+m, [X in vm
o] and Do are genuine
propositions equivalent to X and Do respectively. Hence, it is intuitively rational at tn+m
for B to defer to herself at tn in the suggested way. It follows that SSBR is true.
So far, I have argued that (i) since an agent at tn+m will usually regard herself at tn
as an expert only lacking the information of what sequence of observations she would
make, SSBR is the correct method of deference to her past self, (ii) under the usual
presupposition of perfect memory, SSR follows from SSBR, and (iii) since SSR entails
SSJC, SSJC and SSSC are the correct rules for de nunc updating.
However, I suspect that the reasoning provided in this section doesn’t justify all
instances of SSSC or SSJC. In the next section, I will apply SSSC to the SB problem and
81818181
show that it yields an inconsistent result. In Section G, I will try to identify the source of
the problem and how to fix SSSC and SSJC.
F. The SB Problem and the Inconsistency of SSJC
In this section, we apply SSSC to the SB problem. Interestingly, it will be shown that if
SB updates her credence in H in accordance with SSSC, it leads to an inconsistent result.
Remember the temporal structure of the SB problem:
(s) SB is put to sleep on Sunday.
(m) SB wakes up on Monday.
(m+) SB is told that it is Monday.
We can think about three credal transitions: (i) the transition from s to m, (ii) that from m
to m+, and (iii) that from s to m+. There are two possible strategies for SB to update from
s to m+: (a) updating from s to m+ all at once or (b) updating from s to m and m to m+
step by step. This raises a concern: What if the results of (a) and (b) do not match?
Unfortunately, this concern is legitimate. First, consider this instance of SSSC,
for the updating from m to m+:
(21) Cm+(H)=Cm(H on Monday/MON on Monday)=Cm(H)<1/2.
This is a correct instance of SSSC because SB is sure at m+ of MON&MON (the
conjunction of her observation proposition and the interval information, which happen to
both be MON in this case). First, [MON on Monday] is redundant because “it’s Monday”
82828282
is trivially true on Monday. Second, [H on Monday] is equivalent to H because H is a
genuine proposition whose truth-value is insensitive to time. Thus, Cm+(H)=Cm(H).
However, in the previous chapter, we already determined that Cm(H)<1/2.
Next, consider this instance of SSSC for the updating from s to m+:
(22) Cm+(H)=Cs(H on Monday/W on Monday & MON on Monday)=1/2.
This is a legitimate instance of SSSC because SB is sure at m+ that W&MON was
previously true and MON&MON is presently true. In (22), [W on Monday & MON on
Monday] is redundant because SB knew that she would wake up on Monday and “it’s
Monday” would be true on Monday. Also, [H on Monday] is just the same as H. Hence,
Cm+(H)=Cs(H)=1/2.
Obviously, (21) and (22) are mutually inconsistent. Since both are instances of
SSSC, and so of SSJC, this means that they are inconsistent updating rules.
G. A Diagnosis and a Potential Solution
So SSSC (which is a special case of SSJC) is false because it leads to mutually
inconsistent results. Does this mean that we should totally abandon SSSC and SSJC? No.
In this section, I discuss Elga’s view that temporal knowledge is an essential element of
expertise, and I argue that we have to modify SSSC and SSJC slightly so that the
modified rules do not apply to the cases where the agent suffers from temporal ignorance.
Fortunately, this modification enables us to avoid the aforementioned inconsistency.
83838383
Elga (2007) argues that a rational agent doesn’t have to obey Reflection when
she expects herself to suffer from temporal ignorance:40
There is another sort of information loss, a sort associated with losing track of … what time it is.
Information loss of that sort can also lead to violations of Reflection. For example, suppose that
you are waiting for a train. You are only 50% confident that the train will ever arrive, but you are
sure that if it does arrive, it will arrive in exactly one hour. Since you have no watch, when fifty-
five minutes have in fact elapsed you will be unsure whether an hour has elapsed. As a result, at
that time you will have reduced confidence—say, only 40% confidence—that the train will arrive.
So at the start, you can be sure that when fifty-five minutes have elapsed, your probability that the
train will ever arrive will have gone down to 40%. So your anticipated imperfect ability to keep
track of time creates a violation of Reflection. (Elga 2007, 482)
Let A be the proposition that the train arrives at some time, let CINIT be the agent’s
credence function at the initial moment, and let C55 MIN+ be her credence function in fifty-
five minutes. In the above example, it is an instance of Reflection that CINIT(A/C55 MIN+(
A)=0.4)=0.4, but CINIT(A)=CINIT(A/C55 MIN+(A)=0.4)=0.5. Elga claims that this violation is
understandable because Reflection is a special case of Gaifman’s Expert Principle and the
agent, at the initial time, must not regard herself in fifty-five minutes as an expert given
the expected loss of the track of what time it is.
I find Elga’s claim to be plausible. After all, the agent initially knows that in
fifty-five minutes, only fifty-five minutes will have passed since the initial moment, but,
40 Reflection is the principle that for any proposition X and any real number r, Cn(X/Cn+m(X)=r)=r if defined,
where Cn and Cn+m are agent B’s credence functions at tn and tn+m. Also see the previous chapter for its
relation to Gaifman’s Expert.
84848484
when fifty-five minutes have actually passed, the agent cannot know it because of her
temporal ignorance. Thus, there seems to be some information that she can access
initially but that she cannot access in fifty-five minutes. This relative ignorance makes it
irrational for the agent’s initial self to defer to her future self in fifty-five minutes as
recommended by Reflection.
If temporal knowledge is an essential element of expertise required by Reflection,
it is plausible that temporal knowledge is also a necessary condition of expertise required
by Shifted Backward Reflection. To see this point more clearly, think about the following
variant of the SB problem: SB problem 2. On Sunday, SB knows that she will go
through the following experiment: On that night, evil experimenters will put her to sleep
and toss a fair coin. Case 1: (H) The coin lands heads. Then, she is awakened once in a
room with a big electronic calendar. The calendar is slightly faulty because it has a 0.2
chance of showing the day of tomorrow. Case 2: (T) The coin lands tails. In this case, SB
is awakened in the same room twice, the first time on Monday and the second time on
Tuesday. Plus, the experimenters inject her with a drug with the effect of erasing her
memory of Monday at some time between her two awakenings. In either case, one minute
after she wakes up on Monday or Tuesday, a completely reliable person tells her what
day it is.
As before, let s be the last conscious moment on Sunday, let m be the moment of
wakeup on Monday, and let m+ be the moment of being told that it’s Monday. Also, let
WTUE be the tensed proposition expressed by “SB wakes up today watching the calendar
displaying ‘TODAY IS TUESDAY.’” As a matter of fact, when she wakes up on
Monday, SB receives WTUE as evidence.
85858585
Let’s think about SB’s (i) credal transition from s to m, (ii) that from m to m+,
and (iii) that from s to m+. If we apply SSJC to (i)-(iii), what will it say about her
credences at m and m+ in H?
First, we get the following result by applying SJC (which is a special case of
SSJC) to SB’s credal transition from s to m:
(23) Cm(H)=Cs(H on Monday/WTUE on Monday)Cm(WTUE&MON)+
Cs(H on Tuesday/WTUE on Tuesday)Cm(WTUE&TUE).
Which message the calendar shows on Monday is clearly irrelevant to the result of the
coin toss, and so the first conditional credence is 1/2. Whatever the calendar displays on
Tuesday, waking up on Tuesday entails the coin’s landing tails, and so the second
conditional credence is 0. Hence,
(24) Cm(H)=1/2Cm(WTUE&MON).
Since SB is sure at m that she is waking up reading “TODAY IS TUESDAY” on the
calendar,
(25) Cm(H)=1/2Cm(MON/WTUE).
However, the calendar has a 0.2 chance of displaying the day of tomorrow. Thus, if it is
displaying “TODAY IS TUESDAY,” it has a 0.8 chance of displaying today’s date, in
86868686
which case today is of course Tuesday, but it also has a 0.2 chance of displaying
tomorrow’s date, in which case today is Monday. Intuitively, her credence at m in MON
is 0.2 given WTUE as evidence. Therefore,
(26) Cm(H)=0.1.
Second, we get the following result by applying SSC (which is a special case of
SSSC and therefore of SSJC where the length of the sequence of evidence is 1) to SB’s
credal transition from m to m+:
(27) Cm+(H)=Cm(H on Monday/MON on Monday)=Cm(H).
For SB is told at m+ that it is Monday, but it is no secret that MON is certainly true on
Monday.
Third, we get this result by applying SSSC (which is a special case of SSJC) to
SB’s credal transition from s to m+:
(28) Cm+(H)=Cs(H on Monday/MON on Monday &WTUE on Monday)
=Cs(H)=1/2.
For she knew at s that MON would be true on Monday and, whether the electronic
calendar works correctly or not, it is irrelevant to whether the coin lands on heads.
87878787
Of course, (26)-(28) are jointly inconsistent; if they are all true, SB’s credence in
H is both 0.1 and 1/2 at the same time. Thus, just as in SB problem 1, SSJC leads to
inconsistency in SB problem 2. The difference is that in the latter case, it is easier to see
which result of its application is wrong for what reason: (27) is false because SB doesn’t
know at m what time it is and SSJC doesn’t work correctly when it is used for the credal
transition from a moment of temporal ignorance.
To understand why, remember that I argued for SSJC on the basis of SSBR. For
my present purpose, it is easier to talk directly in terms of SSBR. Hence, see these
instances of SSBR for SB problem 2:
(29) Cm(H/WTUE&MON&Cs(H on Monday/WTUE on Monday)=1/2)=1/2.
(30) Cm(H/WTUE&TUE&Cs(H on Tuesday/WTUE on Tuesday)=0)=0.
(31) Cm+(H/MON&MON&Cm(H on Monday/MON on Monday)=r)=r,
where r=Cm(H).
(32) Cm+(H/[WTUE at prev & MON at pres]&[MON at prev & MON at
pres] & Cs(H on Monday/MON on Monday & WTUE on Monday)=1/2)=1/2.
First, (29) and (30) are the instances of SBR (and therefore of SSBR) for SB’s credal
transition from s to m. Under the assumption that she always remembers her past
credence functions with perfect correctness and confidence, it follows from them that
Cm(H/WTUE&MON)=1/2 and Cm(H/WTUE&TUE)=0, which eventually leads to (26).
Second, (31) is an instance of SBR for her updating from m to m+. Under the same
88888888
assumption, it leads to (27). Third, similarly, (32) leads to (28). Therefore, there is a
contradiction.
The common idea behind these instances of SSBR is that SB must consider
herself at a past moment as an expert in a limited sense.41
For while the agent has no
choice but to depend on her past self to set her present credences, she is also aware that
she has acquired some new information. Thus, I suggested that, roughly, SB must defer
not to her unconditional past credence distribution but to her past credence distribution
conditioned upon the observations that she has made after she has the past credence
distribution.
However, this idea fails to support (31). For when she is told that it is Monday,
SB realizes that it was due to the faultiness of the electronic calendar that she was
strongly biased to the possibility that it was Tuesday then. Being aware of this fact, she
shouldn’t trust her own previous credal judgment, which was based upon that faulty
temporal information.
In general, just as a rational agent cannot consider her future self to be an expert
when she expects to lose track of time in the future, a rational agent cannot regard her
past self to be an expert when she remembers that she lost track of time in the past. For
this reason, I believe that SSBR is not true of cases in which the agent previously did not
know what time it was. Since SSJC gains its plausibility from SSBR, we should apply
SSJC only to cases in which the agent knew what time it was.
Given this diagnosis, let’s think about how it affects the three credal transitions
in the original SB problem: (i) the credal transition from s to m, (ii) that from m to m+,
41 Elga uses “guru” to refer to an agent who is an expert in this limited sense. See Elga (2007).
89898989
and (iii) that from s to m+. If we apply SSJC to all these credal transitions, we are
confronted with a contradiction: By applying SJC to (i), I have shown that SB’s credence
at m in H is less than 1/2. By applying SSC to (ii), I have proven that her credence at m+
in H has the same value as her credence at m in H. However, by applying SSJC to (iii), I
have shown that her credence at m+ in H is 1/2. Hence, 1/2>Cm(H)=Cm+(H)=1/2. A
contradiction.
Fortunately, our discussion in this section suggests that it is faulty to apply SSC
to (ii). For look at Figure 9: As you see in this figure, SB didn’t know at m that it was
Monday, having lost track of what time it was. Thus, at m+, SB cannot regard herself at
m as an expert. If SSJC and its sub-principles are correct only when the agent knew what
time it was at the time from which she is updating, this means that SSC does not apply
correctly to SB’s credal transition from m to m+.
Figure 9: Deference and Temporal Ignorance. It is not necessarily the case that Cm+(H)=Cm(H)=r<1/2.
For it is irrational at m+ for SB to defer to her previous credal judgment, provided that she suffered
from temporal ignorance at m.
90909090
This suggests the following view: When she wakes up on Monday, SB’s
credence in H is lower than 1/2 by SJC applied to (i), and, when she is told that it is
Monday, her credence in H is 1/2 by SSSC applied to (iii). This is exactly the popular
thesis of Thirders.
The discussion until now suggests that we have to modify SSSC and SSJC by
adding the proviso that to update one’s credences from tn to tn+m using these rules, an
agent B must be sure at tn+m that she didn’t suffer from temporal ignorance at tn. Here are
the results of this modification, which I call “Restricted Sequential Shifted Jeffrey
Conditionalization*” and “Restricted Sequential Shifted Strict Conditionalization*”: Let
Eo&Voo∈O be B’s sequential time-observation partition from tn to tn+m, where
Eo=&1≤k≤m(Eko at prevm-k) and Vo=&1≤k≤m(V
ko at prevm-k). Then,
(RSSJC*) Cn+m(X)=Cn(X in vm
o/Do)Cn+m(Eo&Vo) if Eo&Voo∈O is logically optimal
for X and B was free from temporal ignorance at tn,
where Do is the sequential de-indexicalization of Eo under Vo for each o∈O. Next, let
E&V be B’s sequential time-observation partition from tn to tn+m, where E=&1≤k≤m(Ek
at prevm-k) and V=&1≤k≤m(Vk at prevm-k). Then,
(RSSSC*) Cn+m(X)=Cn(X in vm/D)Cn+m(E&V) if E&V is logically optimal for X and
B was free from temporal ignorance at tn,
91919191
where D is the sequential de-indexicalization of E under V. Now, (21) and (27), the
troublesome instances of SSSC in SB problems 1&2, don’t follow from RSSSC*
because its proviso is violated.
Also note that SSC and SJC are the special cases of SSSC and SSJC, respectively,
where the length of sequential timed evidence is 1. Just as we modified the sequential
principles into their restricted versions, we can restrict SSC and SJC with the additional
proviso that the agent was previously free from temporal ignorance. Let’s call the
restricted versions of SSC and SJC “Restricted Shifted Strict Conditionalization*” and
“Restricted Shifted Jeffrey Conditionalization*”. (Since the formal modification is
obvious, I do not provide explicit formulation of RSSC* and RSJC* here.)
In this section, I have provided a diagnosis of the problem discussed in Section F
and discussed how to modify SSSC, SSJC, SSC, and SJC accordingly. The modification
seems successful in removing the aforementioned problem.
H. Conclusion
In this chapter, I have argued for two new principles for sequential updating. I tried to
defend those principles, SSSC and SSJC, by a reasoning similar to that which I used for
SSC and SJC in Chapter II.
However, SSSC and SSJC turned out to be inconsistent when applied to the three
updating paths in the SB problem. With the help of Elga’s discussion, I provided a
diagnosis for the problem of SSSC and SSJC, demonstrating that they are inconsistent
when updating from a credence function when the agent suffers from temporal ignorance.
Accordingly, I suggested weaker modifications of these principles, RSSSC* and RSSJC*.
92929292
These yet new principles are not obviously inconsistent because they don’t apply to cases
in which the agent suffers from temporal ignorance.
Seemingly, this leads to a happy ending: We have updating rules that are free
from obvious inconsistencies and that are justified by plausible arguments. Nevertheless,
I am not perfectly satisfied. On the one hand, the provisos of RSSSC* and RSSJC* are
too restrictive. For it is clearly desirable to have updating rules applicable even to a credal
transition at the initial time of which the agent did suffer from temporal ignorance. On the
other hand, those provisos may not be sufficiently restrictive. For although RSSSC* and
RSSJC* didn’t result in any apparent contradictions in the discussed counterexamples to
SSSC and SSJC, there is no guarantee that we will not find other counterexamples in
which the restricted versions also lead to contradictions.
In the succeeding chapters, I will pursue the following goals: First, I will
formulate and defend general principles for updating de nunc credences that are
applicable to a credal transition with initial temporal ignorance. Second, I will show that
those new principles are free from the problems of SSSC and SSJC that I have discussed
in this chapter.
93939393
CHAPTER IV
UPDATING WITH DE PRIORI INFORMATION
A. Introduction
In the last chapter, I suggested a rule that a rational agent can use to update her de nunc
credences after making a sequence of observations. Now, we face a new challenge:
Suppose that you learn new information about what time it was at an earlier epistemic
moment. For example, you wake up without knowing whether today is Monday or
Tuesday and, after a while, you newly learn that it was Monday when you woke up earlier.
In such a case, what will the rational rule be for updating your credence in a tensed
proposition?
None of the rules discussed in Chapter III will help you to find the answer: First,
SSJC does not apply correctly to any such case. For you can learn new information about
what time it was at an earlier time t only when you did not know “what time it is now” at
t, and I already discussed the fact that it would be irrational for you to use the rule of
SSJC if you are updating from a moment of temporal ignorance. Second, the less general
rules discussed in the last section—SSC, SJC, and SSSC—will not apply correctly to
such a case because they are the sub-principles of SSJC.42
Due to this problem, we need yet another rule for updating. In this chapter, I will
suggest that when an agent newly learns information about what time it was at an earlier
42 Moreover, RSSJC* and its sub-principles (RSSSC*, RSSC*, and RSJC*) do not apply to such a case
because freedom from temporal ignorance is the common proviso of those rules.
94949494
epistemic moment (hereafter: de priori information), the agent ought to update her de
nunc credences by following the updating rule that I call “General Shifted Jeffrey
Conditionalization” (hereafter: GSJC). In particular, I want to achieve these goals: to
defend GSJC and then to illustrate how it works with examples. Finally, I will apply it to
SB’s credal updating from m to m+, arguing that the resulting credence at m+ in H is 1/2.
B. Strategy
In this chapter, my goal is to find a rule for updating de nunc credences. The rule should
be applicable to a credal transition from tn to tn+m even if the agent is ignorant at tn of
what time it is then. To find a clue for such a principle, consider SB’s credal transition
from m to m+. If SB updates her credences in accordance with SSC,
(1) Cm+(H)=Cm(H/MON on Monday)=Cm(H)<1/2.
Here is a rationale for this claim: When told that it is Monday, SB has no choice but to
consult her previous credal judgments to set her present credence in H. However, she
made those judgments before learning that it is Monday. Hence, SB needs to consult her
previous credal judgments conditioned upon MON or something equivalent. In a similar
case, it is generally better to consult one’s previous credal judgments conditioned upon
the de-indexicalization of one’s present observations than to consult those conditioned
upon the present observations themselves.43
Thus, her credence at m+ in H will be equal
to her credence at m in H conditioned upon [MON on Monday].
43 This is because of the so-called outdated conditional credence problem, which I discussed in the last
chapter.
95959595
In the last chapter, I criticized this rationale: The above rationale assumes that
when she is told that it is Monday, SB can regard her previous self as an expert only
lacking what she has just learned. However, her previous self cannot be relied upon as an
expert even in this limited sense. For she did not previously know what day it was then,
and such knowledge is a necessary condition even for the sort of expertise we are talking
about.
To find a better way to set her credence at m+ in H, we must remember that
when SB wakes up on Monday, she is ignorant of what time it is then, and when she is
told that it is Monday, she comes to know something which frees her from her previous
ignorance. With this in mind, I pose three questions: First, what are the contents of SB’s
ignorance and knowledge? Second, what is the logical relation between those contents?
Third, given this relation, what is the rational way for SB to update her credence in H
from m to m+?
Let’s focus on the first question. Waking up on Monday, SB is ignorant of the
fact that “it is Monday at the present moment,” and when she is told that it is Monday,
she acquires the knowledge that “it was Monday at the previous moment” by inference.
The contents of the ignorance and knowledge are those expressed by the sentences
quoted in the last sentence.
This answer is very plausible. For both “the present moment” and “the previous
moment” refer to the same moment (=m), and so it is easy to see how the later
acquirement of knowledge removes the earlier temporal ignorance. Thus we can say:
When she is told that it is Monday, SB is aware of this fact:
96969696
(2) By being ignorant of MON, I previously suffered from temporal ignorance,
from which I have been freed by coming to know [MON at prev].
In general, if an agent previously suffered from ignorance of V, she can be freed from that
ignorance by learning [V at prev] now.44
Now let’s focus on the second question. As previously mentioned, when SB
wakes up, she is ignorant of the fact that “it is presently Monday,” but when told that it is
Monday, she comes to know that “it was previously Monday.” Between the contents of
these ignorance and knowledge, the following logical relation holds: When told that it is
Monday, SB knows that
(3) [MON at prev] is presently true iff MON was previously true.
In general, [V at prev] is presently true iff V was previously true, where V is a tensed
proposition specifying what time it is.
Turning to the third question, I claim that the following equation describes a
rational way for SB to update her credence in H from m to m+:
(4) Cm+(H)=Cm+(H/MON at prev)=Cm(H/MON)=1/2.
44 It is true at m+ that (*) SB is freed from her previous ignorance of MON by coming to know MON.
However, it is not generally true that if an agent previously suffered from the ignorance of V, she can be
freed from that temporal ignorance by learning V. Consider this case: Jake goes to bed not knowing THU,
but he is freed from that ignorance by coming to know [THU at prev] the next morning. In the last sentence,
you cannot replace [THU at prev] with THU salva veritate.
97979797
Here is the rationale for this claim: When she is told that it is Monday, SB realizes that it
was previously Monday but remembers that she did not know what day it was then.
Hence, she will regard her previous self as ignorant of what day it was then. Still, she has
no choice but to depend upon her previous credal judgments to set her present credence
in H. In this situation, she will be better off consulting her previous credal judgments
conditioned upon [MON at prev] or something equivalent. According to (3), [MON at
prev] is true iff MON was true at the previous moment. So it seems rational for her to set
her present credence in H by consulting her previous credence function conditioned upon
MON.
Is this rationale for (4) vulnerable to the same criticism I raised against the
rationale for (1)? No. As (2) says, MON is the content of SB’s temporal ignorance at the
moment of her waking up. Since the consulted conditional credence function is
conditioned upon MON, it is a credal judgment based upon the temporal information that
was correct at the mentioned moment.
Hence, I believe that (4) captures the correct updating pattern for SB’s credal
transition from m to m+. The remaining job is to incorporate this pattern into a new
principle for de nunc updating. For this project, I will proceed in the following order: In
Section C, I will present a general rule for updating de nunc credences. In Sections D and
E, I will defend that rule using a yet new variant of the Conditional Expert Principle. In
Section F, I shall reformulate the thus defended rule into a more readily usable form. In
Section G, I will apply it to SB’s credal transitions from s to m, from m to m+, and from s
to m+. The results will turn out to be mutually consistent. In Section H, I will discuss the
98989898
relation between the rules presented in this chapter and those presented in the earlier
chapters, such as SSC, SJC, SSSC, and SSJC.
C. Updating with De Priori Information
I will begin this section by defining a few important notions related to information about
what time it was. Next, I will present a new principle for de nunc updating, applicable to
a credal transition from a moment of temporal ignorance, and illustrate how it works with
an example. Finally, I will provide a shorter formulation of that principle, which I will
use in later discussions.
To begin, I ask the following questions: Consider an agent B’s credal transition
from tn to tn+m. First, at tn+m, how can B specify what times it had been until an earlier
epistemic moment? Second, how can we (sort of) translate such information from the
context of tn+m to the context of tn? Third, what will the logical relation be between the
original and translated pieces of information?
To answer the first question, I introduce the following definition:
(5) Tensed proposition F is de priori information iff F is [W at prevk] for
some special tensed proposition W and some number k≥1.
For example, consider this case: Example 1. On Sunday night, Jane knows that it is
Saturday or Sunday but does not know which. The next morning, her sincere friend Jeff
tells her, “The last time you were awake it was Sunday.” From this testimony, Jane learns
[SUN at prev(1)]. By definition, it is de priori information.
99999999
This notion of de priori information can be further generalized:
(6) Tensed proposition W is sequential de priori information iff W=(W1 at
prevk)&(W2 at prevk+1)&... &(W
n at prevk+n-1).
For instance, consider this case: Example 2. Briefly waking up on Sunday night, Jane
knows that either [it is Sunday now and it was Saturday at the previous moment] or [it is
Saturday now and it was Friday at the previous moment], but she does not know which.
The next morning, Jeff tells her, “The last time you were awake it was Sunday, and when
you were awake before that moment it was Saturday.” From this testimony, she learns
(SUN at prev1)&(SAT at prev2). By definition, it is sequential de priori information. As
you can easily see, such information is about what times it had been until an earlier
epistemic moment.
Next, I introduce a definition for a special case of sequential de priori
information, namely, information about what times it had ever been until an earlier
epistemic moment: For simplicity, I suppose that B has the first epistemic moment t0, i.e.,
B received her first evidence at t0. (It is possible but less elegant to discuss my view
without this supposition.) Then,
(7) W is a temporal description at tn+m of the epistemic moments until tn iff
W=(W1 at prevm)&(W
2 at prevm+1)&…&(Wn+1
at prevm+n).
100100100100
At tn+m, “prevm+n” refers to t0. Hence, the temporal description at tn+m of the epistemic
moments until tn exhaustively specifies what times it had been at the epistemic moments
between t0 and tn. This answers the first question.
To answer the second question, I introduce this notion:
(8) [W at prevk-m] is the re-indexicalization of [W at prevk] for the m epistemic
moments earlier time.
Consider Example 1 again. When Jane learns [SUN at prev1] from Jeff, SUN (or [SUN at
prev0]) is its re-indexicalization for the previous epistemic moment. We can generalize
this notion for sequential de priori information:
(9) R is the sequential re-indexicalization of W for the m epistemic moments
earlier time iff
(i) W=(W1 at prevk)&(W
2 at prevk+1)&…&(W
n at prevk+n) and
(ii) R=(W1 at prevk-m)&(W
2 at prevk+1-m)&…&(W
n at prevk+n-m).
To understand this definition, consider Example 2 again. When Jane learns (SUN at
prev1)&(SAT at prev2) from Jeff, (SUN at prev0)&(SAT at prev1) is its sequential re-
indexicalization for the previous moment. In a good sense, (SUN at prev0)& (SAT at prev1)
is the “translation” of (SUN at prev1)&(SAT at prev2) from Monday morning to Sunday
night.
101101101101
Now, remember that the temporal description at tn+m of the epistemic moments
until tn exhaustively specifies what times it had been until tn. By definition, such a
temporal description is sequential de priori information, and so it will have a suitable
sequential re-indexicalization. Thus, I make this claim: Consider the temporal description
W at tn+m of the epistemic moments until tn. Then, W’s sequential re-indexicalization R
for the m epistemic moments earlier time will be, in a good sense, the translation of W
from tn+m to tn.
To explain why, we need to answer the third question: Let W be a temporal
description at tn+m of the epistemic moments until tn, and let R be its re-indexicalization
for the m epistemic moments earlier time. Then,
(10) W is true at tn+m iff R was true at tn.
Hence, both W and R describe what times it had been until tn, the former from the point
of view at tn+m and the latter from the point of view at tn.
I have answered all three questions, but I still need one more definition to
precisely formulate the rules I am working toward: Let R and R* be sequential de priori
information such that R=(W1 at prevk)&(W
2 at prevk+1)&... &(W
n at prevk+n-1) and
R*=(W*1 at prevk)&(W*
2 at prevk+1)&... &(W*
n at prevk+n-1). In a good sense, R ascribes
w1 to the k epistemic moments earlier time, w
2 to the k+1 epistemic moments earlier
time, …, wn+1
to the k+n-1 epistemic moments earlier time; similarly for R*. Then,
102102102102
(11) R* is better de priori information than R iff for any k∈1,...,m,
w*k⊆w
k, and for some k∈1,...,m, w*
k⊂wk.
In other words, R* is better than R exactly when R* attributes at least equally narrow
intervals to the mentioned epistemic moments and R* attributes a strictly narrower
interval to one of the epistemic moments than R. Let R be any sequential de priori
information. Then, for any probability function C and tensed propositions X and Y,
(12) R is well-specified de priori information with respect to <C, X, Y>
iff C(X/Y&R)=C(X/Y&R*) for any sequential de priori information R*
better than R.
When this condition is satisfied, I will often say informally that C(X/Y&R) is conditioned
upon well-specified de priori information. This definition allows us to formulate the
desired rules without precisely specifying what times it had been until the moment from
which updating occurs.
Now, I am ready to formulate my first updating rule in this chapter, “General
Shifted Strict Conditionalization”: Consider a sequence of observations E1, E
2, ...E
m, a
sequence of intervals v1, v
2, ... v
m, and another sequence of intervals w
1, w
2, ... w
n+1.
Assume that (a) the truth-value of X is invariant within vm
and that of Ek is invariant
within vk for each k∈1,...,m and (b) Cn(X in v
m/(E
1 in v
1)&... &(E
m in v
m)&(W
1 at
prevm)&...& (Wn+1
at prevm+n)) is conditioned upon a well-specified temporal description.
Then, for any tensed proposition X,
103103103103
(GSSC) Cn+m(X)=Cn(X in vm
/(E1 in v
1)&...&(E
m in v
m)&(W
1 at prevm)&...& (W
n+1
at prevm+n)) if
(i) B is sure at tn+m that for each k∈1,...,m, [Ek was/is true and it was/is
vk] m-k epistemic moments ago, and
(ii) B is sure at tn+m that for any k∈1,...,n+1, [it was wk] m+k-1 epistemic
moments ago,
where Cn and Cn+m are B’s credence functions at tn and tn+m. Less formally, we can
rewrite SSSC in this way: Let E be (E1 at prevm-1)&...&(E
m at prev0), V be (V
1 at prevm-1)
&...&(Vm
at prev0), and W be (W1 at prevm)&(W
2 at prevm+1) &…&(Wn+1
at prevm+n).
Assume that (a) and (b) are true. Then,
(GSSC) Cn+m(X)=Cn(the de-indexicalization of X under Vm
/the sequential de-
indexicalization of E under V & the sequential re-indexicalization of W for
the m epistemic moments earlier time) if B has certainly learned until tn+m
that E&V&W is true.
Unfortunately, GSSC is a principle with an extremely narrow range of
application. In order to use it, we need to know what time it has been since our very first
observation. Hence, we are forced to move to our next, more general updating principle.
First, let me outline the core idea: Let Eo&Vo&Woo∈O be a partition such that Eo=(E1
o at
prevm-1)&...& (Em
o at prev0), Vo=(V1
o at prevm-1)&...&(Vm
o at prev0), and Wo=(W1
o at
104104104104
prevm)&...& (Wn+1
o at prev0), for each o∈O. I consider each Eo&Vo to represent a possible
evidential scenario and each Wo to represent what times it had been until tn. For
simplicity, let O be 1, 2, ...p. By GSSC:
Cn+m(X) would be
Cn(X in vm
1/D1&R1) if B were sure at tn+m of E1&V1&W1,
Cn(X in vm
2/D2&R2) if B were sure at tn+m of E2&V2&W2,
...
Cn(X in vm
p/Dp&Rp) if B were sure at tn+m of Ep&Vp&Wp,
where Do is the sequential de-indexicalization of Eo under Vo and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time for each o∈O. If we
accept GSSC (at least for its narrow range of application), it is natural that B’s credence
at tn+m in X is the weighted average of values on the right-hand sides of the above
equations with the weights coming from B’s credences at tn+m in Eo&Vo&Wo.
To implement this idea, we need to finish some formal homework first: Consider
a partition &1≤k≤m((Eko&V
ko) at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1)o∈O such that (i)
Cn+m(&1≤k≤m((Eko&V
ko) at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1))>0 for each o∈O and (ii)
∑o∈OCn+m(&1≤k≤m((Eko&V
ko) at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1))=1 where Cn+m is an
agent B’s credence function at tn+m. (The intended interpretation of this partition is that
each member is a hypothesis about (i) what observations have been made at what times
after tn and (ii) what times it had been until tn.) I will call any member of this partition
105105105105
“(B’s) general time-observation proposition from tn to tn+m.” If that partition also satisfies
the condition that for each o∈O, Cn(&1≤k≤m(Eko in v
ko)&&1≤k≤n+1(W
ko at prevk-1))>0,
then I will call the partition “(B’s) general time-observation partition from tn to tn+m.”
Now, I am ready to present my next updating rule, called “General Shifted
Jeffrey Conditionalization”: Let &1≤k≤m((Eko&V
ko) at prevm-k) & &1≤k≤n+1(W
ko at
prevm+k-1)o∈O be B’s sequential time-observation partition from tn to tn+m over [t0,tn+m].
Assume that for any o∈O (a) the truth-value of X is invariant within vm
o and that of Eko is
invariant within each vko for any k∈1,...,m and (b) Cn+m(&1≤k≤m((E
ko&V
ko) at prevm-
k)&&1≤k≤n+1(Wko at prevm+k-1)) is conditioned upon a well-specified temporal description.
Then,
(GSJC) Cn+m(X)=Σo∈O[Cn(X in vm
o/&1≤k≤m(Eko in v
ko)&&1≤k≤n+1(W
ko at prevk-1))
Cn+m(&1≤k≤m((Eko&V
ko) at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1))],
where Cn and Cn+m are B’s credence functions at tn and tn+m. Less formally: Let
Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and Wo=&1≤k≤n+1(W
ko at prevm+k-1)
for each o∈O, so that Eo&Vo&Woo∈O is the same as &1≤k≤m((Eko&V
ko) at prevm-
k)&&1≤k≤n+1(Wko at prevm+k-1)o∈O, B’s general time-observation partition from tn to tn+m.
Assume that for each o∈O (a) and (b) are true. Then,
106106106106
(GSJC) Cn+m(X) = the weighted average of Cn(the de-indexicalization of X under
Vm
o/the sequential de-indexicalization of Eo under Vo & the sequential re-
indexicalization of Wo) with the weights coming from Cn+m(Eo&Vo&Wo)
where Cn and Cn+m are B’s credence functions at tn and tn+m. I think this is a natural
generalization of GSSC for when the agent is not sure of what sequence of evidence she
has received at what times and what times it had been before receiving any sequence of
evidence in consideration.
To see how GSSC and GSJC work, consider the following example: Example 3.
Let R be the tensed proposition expressed by “it rains today in Boston,” and let P be that
expressed by “there is a form of precipitation today in Boston.” Jane is born on Saturday,
knowing that (SAT) it is Saturday.
Just after her birth, she falls asleep and then wakes up on Sunday. She learns that
(SUN) it is Sunday. At this time, Jane assigns the credence of 0.8 to [R on Monday] given
[P on Monday]. Immediately after waking up, she takes a sleeping pill that will make her
wake up either on Monday or Tuesday, but she will not know which day it is when she
wakes up.
In fact, Jane wakes up on Monday. On being awakened, she is told that there is a
form of precipitation today. For the mentioned reason, she does not know whether (MON)
it is Monday or (TUE) it is Tuesday; indeed, she assigns the credence of 0.5 to each of
MON and TUE. (Hence, 0.5 is her credence at this moment in P&MON.) Later on that
day, she is told that it is Monday.
107107107107
For brevity, let b be the moment of Jane’s birth on Saturday, let s be the moment
of waking up on Sunday, let m be the moment of waking up on Monday, and let m+ be
the moment of being told that it is Monday. We assume that she does not observe
anything between these moments. What, then, is her rational credence at m+ in R?
First, let’s focus on her credal transition from m to m+. By GSSC,
(13) Cm+(R)=Cm(R on Monday/(MON on Monday)&(MON at
prev0)&(SUN at prev1)&(SAT at prev2)).
The conditional part of the right-hand side was acquired as in Figure 10:
We can simplify (13) by using these facts: Jane was sure at m that MON is true on
Monday, it was Sunday at the last epistemic moment (=s), and it was Saturday two
epistemic moments ago (=b). At m, she believed MON to the degree of 1/2. Hence,
(14) Cm+(R)=Cm(R on Monday/MON)=2Cm((R on Monday)&MON).
Time b s m m+
An observation at m MON
Temp. desc. until m SAT at prev2 SUN at prev1 MON at prev0
SAT at prev2 SUN at prev1 MON at prev0MON on Monday
Sequential re-indexicalization
de-
indexicali
zations
Figure 10: Reindexicalization and Deindexicalization 1. The sequential de
priori information (MON at prev1)&(SUN at prev2)&(SAT at prev3) is re-
indexicalized to (MON at prev0)&(SUN at prev1)&(SAT at prev2) and the
108108108108
Let R*=(R on Monday)&MON; thus, Cm+(R)=2Cm(R*). To use this equation, we need to
find her credence at m in R*.
Focus on Jane’s credal transition from s to m: She is not sure at m whether it is
Monday or Tuesday, but she remembers that it was Sunday at the previous moment and
that it was Saturday two moments ago. By GSJC,
(15) Cm(R*)=
Cs(R* on Monday/(P on Monday)&(SUN at prev0)&(SAT at prev1))
Cm((P&MON)&(SUN at prev1)&(SAT at prev2) )+
Cs(R* on Tuesday/(P on Tuesday)&(SUN at prev0)&(SAT at prev1))
Cm((P&TUE)&(SUN at prev1)&(SAT at prev2)).
For [P on Monday] is the de-indexicalization of P under MON, and [P on Tuesday] is
that of P under MON, and (SUN at prev0)&(SAT at prev1) is the re-indexicalization of
(SUN at prev1)&(SAT at prev2) for the previous epistemic moment, as in Figure 11. Also,
observe these facts: (i) [R* on Monday] is equivalent to [R on Monday].45
(ii) [R* on
Tuesday] is impossible.46
(iii) Jane was sure at s of [SUN at prev0]&[SAT at prev1]. (iv)
Jane is sure at m of [SUN at prev1]&[SAT at prev2]. Hence,
45 For [((R on Monday)&MON) on Monday]=[((R on Monday) on Monday)&(MON on Monday)]=(R on
Monday).
46
For [((R on Monday)&MON) on Tuesday] entails (MON on Tuesday), which is a contradiction.
109109109109
(16) Cm(R*)=Cs(R on Monday/P on Monday)Cm(P&MON)=0.4.
From (14) and (16), it follows that her credence at m+ in R is 0.8. This is an intuitive
result, as her credence at s in raining on Monday was 0.8, conditional on precipitation on
Monday. She later learns that there is a form of precipitation today and that today is
Monday. Hence, it follows that there is a form of precipitation on Monday after all. Given
these facts, it is natural that her credence at m+ (∈ Monday) is 0.8 in R.
Finally, I want to formulate GSJC and GSSC in more succinct forms: Let
Eo&Vo&Woo∈O be an agent B’s general time-observation partition from tn to tn+m,
where Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and Wo=&1≤k≤n+1( W
m+ko at
Time b s m
An observation at m P
Temp. desc. until m SAT at prev2 SUN at prev1
The 1st seq. re-indexicalization
& de-indexicalizationSAT at prev1 SUN at prev0 P on Monday
The 2nd seq. re-indexicalization
& de-indexicalizationSAT at prev1 SUN at prev0 P on Tuesday
Seq. re-indexicalization
de-
indexicali
zations
Figure 11: Reindexicalization and Deindexicalization 2. The observation P is de-
indexicalized to (P on Monday) and (P on Tuesday), and sequential de priori
information (SUN at prev1)&(SAT at prev2) is re-indexicalized to (SUN at
prev0)&(SAT at prev1).
110110110110
prevm+k-1). Given this partition, I introduce these definitions: (i) Eo&Vo&Woo∈O is
probabilistically optimal for X iff for any o∈O, Cn+m(X in vm
o/Do&Ro) is conditioned
upon a well-specified temporal description, where Do=&1≤k≤m(Eko in v
ko) and
Ro=&1≤k≤n+1(Wko at prevk-1). Also, remember the definition of logical optimality in the
last chapter. Given these two definitions, (ii) Eo&Vo&Woo∈O is optimal for X iff
Eo&Voo∈O is logically optimal for X and Eo&Vo&Woo∈O is probabilistically optimal
for X. Then,
(GSJC) Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if Eo&Vo&Woo∈O
is optimal for X,
where Do=&1≤k≤m(Eko in v
ko) and Ro=&1≤k≤n+1(W
ko at prevk-1). Next, let E&V&W be
B’s sequential time-observation partition from tn to tn+m, where E=&1≤k≤m(Ek at prevm-k),
V=&1≤k≤m(Vk at prevm-k) and W=&1≤k≤n+1(W
m+k at prevm+k-1). Then,
(GSSC) Cn+m(X)=Cn(X in vm/D&R) if E&V&W is optimal for X,
where D=&1≤k≤m(Ek in v
k) and R=&1≤k≤n+1(W
k at prevk-1). Obviously, these
reformulations are equivalent to the original.
111111111111
So far, I have presented GSJC and GSSC, illustrated how they work with an
example, and provided shorter formulations for them. The next step is to defend the new
updating principles with a new variant of the Conditional Expert Principle.
D. Temporal Conditional Multiple Expert Principle
In the last chapter, I introduced the Temporal Conditional Expert Principle. That principle
described an epistemic relation between two agents located at different times. As such, it
takes only their times and observations into consideration. In this section, I will present a
new principle, which is similar to TCE but assumes additional agents contributing to the
epistemic cooperation.
First, consider two agents Bn+m, located at time tn+m, and Bn, located at time tn,
where tn+m>tn; for convenience, I will call Bn+m “the client” and Bn “the expert.” As the
names indicate, the client wants to set her credences at tn+m by consulting Bn’s credal
opinion at tn. (I will often omit “at tn+m” and “at tn.”) Here, I suppose that the expert’s
credal judgment is not dependent upon any other agent’s data or judgment. See Figure 12.
Call this type of situation “a two agent situation.” In this situation, what is the rational
way for the client to set her credences by checking the expert’s?
Figure 12: Judgmental Dependence. The arrow line indicates the judgmental
dependence. There is no informational dependence on any other agents.
112112112112
Earlier, I argued for the following answer: For any tensed proposition X,
(TCE) Cn+m(X/E&V &prn(X′/E′)=r)=r if the left-hand side has a defined value
and the following conditions are satisfied:
a) Cn+m is the client’s credence function at tn+m, and prn is the expert’s
credence function at tn.
b) All information had at tn+m by the client is accessible to the expert
at tn, possibly except E, and the expert perhaps does not know at tn
whether E′ is true.
c) X′ and E′ are tensed propositions such that the client knows at tn+m
that if V is true, [X is presently true iff X′ was true at tn] and [E is
presently true iff E′ was true at tn].
I briefly repeat the rationale that I provided for this principle: By (a), it seems rational
that Cn+m restricts Cn given the agent’s intention to consult the expert’s opinion in order
to set her de nunc credences. By (b), it will however be irrational for the client to set her
credence in X to be simply the same as the credence that the expert assigned to X. By (c),
the client will think that if V is true, it is best to assign r to X conditional on E provided
that Ex’s credence in X′ given E′ is r; for, she knows that if V is true, [X is presently true
iff X′ was true when the expert had the consulted credal opinion] and [E is presently true
iff E′ was true when the expert had the consulted credal opinion].
I still believe that TCE makes sense when there are no agents other than the
client and the expert to take into consideration. However, think about the following
113113113113
situation: Consider n+m+1 agents, Bn+m, …, Bn, …, B0, located at different moments,
tn+m, …, tn, …, t0, where tn+m>…>t0; for convenience, we will call Bn+m “the client,”
Bn+m, …, Bn+1 “the direct data providers,” Bn “the expert,” and Bn, …, B0 “the indirect
data providers.” We suppose that the client intends to judge the probability of X with the
help of the expert, given the data observed by the direct data providers. Also, the expert’s
judgment, consulted by the client, is based upon the data observed by the indirect data
providers. See Figure 13. Call this situation “a multiple agent situation.”47
In this
situation, what is the rational way for the client to set her present credence?
As an answer, I suggest a principle that I call “the Temporal Conditional
Multiple Expert Principle”: Let E be a tensed proposition specifying the observations that
47 In their paper (forthcoming), Dietrich and List discuss the topic of how one can aggregate the opinions
of multiple agents in a rational way. I believe that such a theory will be useful also for clarifying how a
single person can aggregate the opinions of her multiple selves in the past in a rational way. Another
important topic, which has been left unstudied as far as I know, is that of how an individual is supposed to
aggregate the de se opinions of multiple agents.
Figure 13: Direct and Indirect Data Providers. Here, Bn+m is the client and Bn is
the expert. The dashed arrow line indicates the informational dependence, while
the solid arrow line—from Bn+m to Bn—indicates the judgmental dependence.
114114114114
have been made by the direct data providers, let V be a tensed proposition about in what
times the direct data providers are located, and let W be a tensed proposition about in
what times the indirect data providers are located. Then, for any tensed proposition X,
(TCME) Cn+m(X/E&V&W&prn(X′/E′&W′)=r)=r, if the left-hand side has a defined
value and the following conditions are satisfied:
(d) Cn+m is the client’s credence function at tn+m and prn is the expert’s
credence function at tn.
(e) All information had at tn+m by the client is accessible to the expert
at tn possibly except E, and so the expert perhaps does not know at
tn whether E′ is true.
(f) X′ and E′ are tensed propositions such that the client knows at tn+m
that if V is true, [X is presently true iff X′ was true at tn] and [E is
presently true iff E′ was true at tn].
(g) W′ is a tensed proposition such that B knows at tn+m that [W is
presently true iff W′ was true at tn].
Here, the main difference is (d): According to TCE, the client does not have to take the
temporal locations of the agents providing data to the expert. According to TCME, the
client needs to take those agents’ temporal locations into consideration; conditional on
the assumption that the indirect data providers were located at the times specified by W,
the client needs to consult the expert’s credence in X′ not only conditioned upon E′, but
115115115115
also upon W′, where W′ specifies the same temporal locations of the indirect data
providers as specified by W.
Why this difference? In a multiple agent situation, the client will be aware that
the expert’s credal opinion was made by depending upon the data from the indirect data
providers. In order for the expert to correctly interpret those data, he will need the
information about when those data were observed; in other words, he will need the
information about the indirect data providers’ temporal locations.
To appreciate this point, let’s consider an analogous example: Example 4. Four
meteorologists are flying in balloons in the New England sky. Let’s call them “B3,”
“B2,”“B1,” and “B0,” in the order of spatial proximity to B3. Suppose that B3 is judging the
probability of rain in her region with the help of the other meteorologists. I assume that
each Bk is equipped with a walkie-talkie but does not communicate with the other
meteorologists unless Bk is contacted by a Bi>k or needs the data or judgment of some Bi<k.
(Hence, the information flows from B0 to B3 but not in the other direction.) In this
situation, what will be the best strategy for B3 to make a credal judgment about rain in her
region?
One good strategy would be for her to “delegate” some required tasks to, say, B1,
so that while B3 gathers data from B3 and B2, she depends upon the judgment of B1. (Note
that this is not to disregard the observations made by B1 and B0 because, if rational, B1
will judge by taking their observations into consideration.) Let’s call B3 “the client,” B3
and B2 “the direct data providers,” B1 “the expert,” and B1 and B0 “the indirect data
providers.”
116116116116
First, the client will need the information about the direct data providers’ spatial
locations in order to correctly judge the probability of rain using their data. This is
because without knowing where the observed events are occurring, it will be difficult to
correctly judge the relevance of the observations made by the direct data providers to the
possibility of rain in her region.48
For example, suppose that E represents the data
observed by the direct data providers, where E=(5°C temperature is being observed by
B3)&(thick clouds are being observed by B2). To correctly interpret E, the client (=B3)
will need information about where the direct data providers observed the conjuncts of E,
such as L=(B3 is located in Amherst)&(B2 is located in Pelham).
Second, the client also will need the spatial locations of the indirect data
providers in order to rationally utilize the expert’s credal judgment. Intuitively, given
E&L, the client will consult the expert’s probability of rain in Amherst conditioned upon
E ′, where E ′=(5°C temperature is being observed in Amherst)& (thick clouds are being
observed in Pelham). Now, if the expert is rational, he will make this credal judgment on
the basis of the data from the indirect data providers. Consequently, the expert will need
the spatial locations of the indirect data providers to correctly interpret the data from
them, just as the client needs the spatial locations of the direct data providers to correctly
interpret their data.
To see this point clearly, suppose that the expert has the information that (F) a
strong wind is observed by B1, and the wind is observed to be blowing from east to west
by B0. Then, compare two possible spatial locations of B1 and B0:
48 Also, she will need to know the region in which she is located to know the region for which she is
judging the probability of rain.
117117117117
(M) B1 is located in Belchertown, and B0 is located in Ware.
(M*) B1 is located in Newton, and B0 is located in Cambridge.
From the point of view in Amherst, a strong east wind in nearby eastern areas such as
Belchertown and Ware will elevate the probability of rain in Amherst given E′, but the
same wind in distant lands like Newton and Cambridge will not be particularly relevant
to the possibility of rain in Amherst on the condition of E′.49 Hence,
(17) prn(raining in Amherst/E ′&M)≠prn(raining in Amherst/E ′&M*),
where prn is the expert’s credence function. Now, assume that the expert was almost sure
of M*, so that
(18) prn(raining in Amherst/E′)≈prn(raining in Amherst/E′&M*),
but also that the client is quite sure of M. In that case, it will be irrational that
(19) Cn+m(raining /E&L &prn(raining in Amherst/E′)=r)=r if defined,
49 For the relevant geographical data, see the entry of “Massachusetts” in http://www.wikipedia.org.
118118118118
where Cn+m is the client’s credence function. For the expert’s above conditional credence
was, from the client’s point of view, a judgment largely based upon wrong information
about the indirect data providers’ spatial locations.
This problem is due to the potential difference between the client’s and the
expert’s opinions about the indirect data providers’ spatial locations. One way to bracket
out this difference is to consult the expert’s credence conditioned upon the (from the
client’s point of view) correct information about the spatial locations of the indirect data
providers. This suggests the following relation between the client’s and the expert’s
credal opinions: For any real number r,
(20) Cn+m(raining/E&L&M&prn(raining in Amherst/E′&M)=r)=r if
defined.
It is not difficult to apply the same idea to a case in which the agents are located at
different times; consider example 5. Four meteorologists have made observations about
the weather of Amherst on different days. Call the meteorologists “B3,” “B2,”“B1,” and
“B0,” in the reverse order of time. We suppose that B3, “the client,” is making a credal
judgment about whether it will rain today with the help of the other three agents. In
particular, she depends upon the data from B3, B2, “the direct data providers,” and the
judgment of B1, “the expert.” It is clear that she also comes to indirectly rely upon the
data from B1, B0, “the indirect data providers.” In this case, what will be the correct way
for the client to make her credal judgment about whether it will rain in Amherst today?
119119119119
First, she will need to know the temporal locations of the direct data providers in
order to correctly interpret the data from them. Let E=(5°C temperature is being observed
by B3)&(thick clouds were observed by B2), and V=(B3 is located in Saturday)&(B2 is
located in Sunday). Given V, the client can process E into the equivalent data E′=(the 5°C
temperature is being observed on Sunday)&(thick clouds were observed on Sunday).
Second, she will need to know the temporal locations of the indirect data
providers in order to correctly utilize the expert’s credal judgment. For the expert’s credal
opinion must have been produced on the bases of their data. Suppose that the expert had
the information that (F) B0 had met somebody telling him that for the next three days,
once thick clouds have formed, they will not go away quickly, and B1 was told that the
guy she met yesterday was a very good expert about cloud forming. Consider these
temporal locations of the indirect data providers:
(W) B1 was located on Saturday, and B0 had been located on Friday.
(W*) B1 was located on Thursday, and B0 had been located on Wednesday.
Clearly,
(21) prn(raining on Monday/E′&W)≠prn(raining on Monday/E′&W*),
where prn is the expert’s credence function. Assuming that the client is almost sure of W,
but the expert is almost sure of W*, it will be irrational that
120120120120
(22) Cn+m(raining /E&V&prn(raining on Monday/E′)=r)=r if defined,
where Cn+m is the client’s credence function. For the expert’s above conditional credence
is largely based on, from the client’s point of view, wrong information about the indirect
data providers’ temporal locations, which was essential to judging correctly E′’s
relevance to whether or not it will rain on Monday. Rather, the rational way that the client
would utilize the expert’s opinion is the following:
(23) Cn+m(raining /E&V&W&prn(raining on Monday/E′&W)=r)=r if
defined.
Until now, we have discussed cases in which the client has a means of specifying
the indirect data providers’ locations in non-indexical ways. However, I do not think the
lesson we have learned from these examples depends upon the existence of such a non-
indexical method of specifying the indirect data providers’ locations. So let W and W′ be
the tensed propositions specifying the indirect data providers’ temporal locations such
that from the client’s point of view, (*) W is presently true iff W′ was true at the expert’s
time. Then, it will be the case that
(24) Cn+m(X/E&V&W&prn(X′/E′&W′)=r)=r if defined,
where X′ and E′ satisfy TCME’s provisos (a)-(c) with respect to X, E, and V. For W tells
the client that W′ was [the tensed proposition specifying the indirect data providers’
121121121121
temporal locations] that was true at the expert’s time. By introducing (*) as the fourth
proviso (d), we acquire the general principle TCME.
In this section, I have argued that we need to modify TCE in order to capture the
rational method for an agent to defer to the credal judgment of another agent located at an
earlier moment, where both agents are provided relevant data.
E. A Defense of GSJC
In this section, I will present a new principle, “General Shifted Sequential Rigidity”
(hereafter: GSR), and defend it by making use of TCME. After defending it, I will point
out that GSR entails GSJC. Since GSSC is a special case of GSJC, the two principles
presented in the last section will have been defended.
I start by presenting GSR: Let Eo&Vo&Woo∈O be B’s general time-observation
partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and
Wo=&1≤k≤n+1(Wko at prevm+k-1). Next, suppose that Eo&Vo&Woo∈O is optimal for X.
Then,
(GSR) Cn+m(X/Eo&Vo&Wo)=Cn(X in vm
o/Do&Ro),
where Do is the sequential de-indexicalization of Eo under Vo and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time. (In other words,
122122122122
Do=&1≤k≤m(Eko in v
ko) and Ro=&1≤k≤n+1(W
ko at prevk-1).) It is not difficult to see that
GSR entails GSJC.50
How do we defend GSR? We defend it by using TCME. First, let Bn+m, …, B0 be
the agent B’s selves at tn+m, …, t0, where tn+m>…>t0. We suppose that Bn+m sets her
credence in X with the help of Bn+m-1, …, B0. One way to do so will be to collect the data
from Bn+m, …, Bn+1 and delegate the job of making a suitable conditional credal judgment
to her past self, say, Bn. Accordingly, let’s call Bn+m “the client” and Bn “the expert.”
One important question is whether this is a two agent situation or a multiple
agent situation. Clearly, this is a case of the latter. Not only does the client depend upon
Bn+m, …, Bn+1 to acquire extra data, but the expert must also depend upon Bn, …, B0 to
make the suitable conditional credal judgment. Accordingly, let’s call Bn+m, …, Bn+1 “the
direct data providers” and Bn, …, B0 “the indirect data providers.”
Once put in this way, it is plausible that TCME applies to this case, since I have
previously argued that in a multiple agent case, TCME, not TCE, is the principle
describing the epistemic relation between the client and the expert. Then, we can derive
the principle that I call the “General Shifted Backward Reflection Principle” from TCME:
Let X be any tensed proposition and Eo&Vo&Woo∈O be B’s general time-observation
partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and
Wo=&1≤k≤n+1(Wko at prevm+k-1). Next, let Dο be the sequential de-indexicalization of Eo
50 Suppose GSR. So Cn+m(X/Eo&Vo&Wo)=Cn(X in v
mo/Do&Ro) for any o∈O. Since Eo&Vo&Woo∈O is a
general time-observation partition from tn to tn+m, (i) ∑o∈OCn+m(X/Eo&Vo&Wo)=1 and (ii) (X/Eo&Vo&Wo)>0
for each o∈O. Thus, Cn+m(X)=∑o∈OCn+m(X&Eo&Vo&Wo)=∑o∈OCn+m(X/Eo&Vo&Wo)Cn+m(Eo&Vo&Wo)= (by
supposition) ∑o∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo). Done.
123123123123
under Vo (i.e., Dο=&1≤k≤m(Eko in v
ko)), and let Ro be the sequential re-indexicalization of
Wο for the m epistemic moments earlier time (i.e., Wo= &1≤k≤n+1(Wko at prevk-1)).
Assume that Eo&Vo&Woo∈O is optimal for X. Then,
(GSBR) Cn+m(X/Eο&Vο&Wο&Cn(X in vm
o/Dο&Rο)=r)=r if defined.
For (a) Cn+m is the client’s credence function, and Cn is the expert’s, (b) Eo specifies the
data accessible to the client but perhaps not to the expert, (c) the client knows (at tn+m)
that X is presently true iff [X in vm
o] was true m epistemic moments ago, and Eo is
presently true iff Do was true m epistemic moments ago, and (d) the client also knows (at
tn+m) that Wo is presently true iff Ro was true m epistemic moments ago. Since the
elements composing GSBR satisfy the provisos of TCME, GSBR is a special case of
TCME. GSR is derivable from GSBR under the presupposition that B always remembers
her past credence function with perfect confidence and correctness.51
Remember that GSBR entails GSR and GSR entails GSJC. Therefore, we have
good reason to accept GSJC.
F. Too Far Past Does Not Matter
The earlier formulation of GSJC has a practical problem: In most credal transitions, we
do not worry about what time it was at t if t is a moment sufficiently far past. Even in
such a case, GSJC asks us to take such a matter into consideration. In this section, I will
51 Let r=Cn(X in vj/Do&Ro). By the presupposition of perfect memory, Cn+m(Cn(X in vj/Do&Ro))=r)=1.
Thus, Cn+m(X/Eo&Vo&Wo)=Cn+m(X/Eo&Vo&Wo&Cn(X in vj/Do&Ro)=r)=(by GSBR)r=Cn(X in vj/Do&Ro).
Done.
124124124124
provide another formulation of GSJC, which is (i) relatively free from this problem and
(ii) still equivalent to the original formulation of GSJC.
First, I review GSJC: Let Eo&Vo&Woo∈O be an agent B’s general time-
observation partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at
prevm-k), and Wo=&1≤k≤n+1(Vko at prevm+k-1). Then,
(GSJC) Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if Eo&Vo&Woo∈O
is optimal for X,
where Do is the sequential de-indexicalization of Eo under Vo (i.e., Do=&1≤k≤m(Eko in v
ko))
and Ro is the sequential re-indexicalization of Wo for the m epistemic moments earlier
time (i.e., Ro=&1≤k≤n+1(Wko at prevk-1)).
We can expect a complaint: If I update in accordance with GSJC, then I have to
presently assign to X the weighted average of [X in vm
o] given Do&Ro (with the weights
coming from…), where Ro specifies what times it had been at all my epistemic moments
until that from which I am updating. These epistemic moments even include the moment
of my first observation. This appears to be absurdly demanding. For example, why must I
take my birth time into consideration, in judging the probability of rain today?
This is a fair complaint. There should be a way in which one can rationally judge
the probability of rain today without worrying about when she was born, when she fell in
love for the first time, etc. So, I reformulate GSJC into a principle that does not mention
125125125125
epistemic moments that are too far in the past to be relevant. First, I suggest a modified
definition of general time-observation partition. The core idea is that first we can remove
from the given general time-observation partition its elements ascribing times to
epistemic moments too far in the past to be relevant, and second, we can formulate the
updating principle that works with the remaining partition.
Here is the first step: Again, let Eo&Vo&Woo∈O be an agent B’s general time-
observation partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at
prevm-k), and Wo=&1≤k≤n+1(Wko at prevm+k-1). Consider some i∈1,...,n+1. Given
Eo&Vo&Woo∈O and i, let Ep&Vp&Wpp∈P be the maximal partition such that for each
p∈P, there exists o∈O such that Ep=Eo, Vp=Vo, and Wp=&1≤k≤n-i(Wko at prevm+k-1).
52
Given these partitions, I will say that Ep&Vp&Wp is the abbreviation of Eo&Vo&Wo. It
follows that for any o∈O, there exists p∈P such that Eo&Vo&Wo=(Ep&Vp&Wp&W*p),
where W*p=&n-i +1≤k≤n+1(Wm+k
o at prevm+k-1). In such a case, I will say that W*p is a
complement of Ep&Vp&Wp for Eo&Vo&Wo. The point is that Ep&Vp&Wpp∈P is the
same as Eo&Vo&Woo∈O except that each Wp may include only a fragment of
corresponding Wo. In such a case, I will say that Ep&Vp&Wpp∈P is (B’s) general time-
observation partition from tn to tn+m over [ti,tn+m].53
Also, I will say that Ep&Vp&Wpp∈P
52 Here, I suppose that if “1≤k≤n-i” is not satisfied by any number k, then &1≤k≤n-i(W
ko at prevm+k-1) is
vacuously true. For example, if i=n+1, &1≤k≤n-i(Wko at prevm+k-1)=&1≤k≤-1(W
ko at prevm+k-1)=T, where T is any
tautology.
53
Note: If i=n+1, then Wp is vacuous (see the previous footnote) and so Ep&Vp&Wpp∈P=Ep&Vpp∈P=
Eo&Voo∈O.
126126126126
is sufficiently inclusive for a tensed proposition X iff for each o∈O and p∈P, if
Ep&Vp&Wp is an abbreviation of Eo&Vo&Wo and W*p is the complement, then Cn(X in
vm
p/Dp&Rp)= Cn(X in vm
p/Dp&Rp&R*p), where Dp is the sequential de-indexicalization
of Ep under Vp and Rp and R*p are the sequential re-indexicalizations of Wp and W*p for
the m epistemic moments earlier time.
Here is the second step: Let Ep&Vp&Wpp∈P be a general time-observation
partition from tn to tn+m over [ti,tn+m]. Then,
(GSJC-) Cn+m(X)=Σp∈PCn(X in v
mp/Dp&Rp)Cn+m(Ep&Vp&Wp) if Ep&Vp&Wpp∈P is
optimal and sufficiently inclusive for X,
where Dp is the sequential de-indexicalization of Ep under Vp, and Rp is the sequential re-
indexicalization of Wp for the m epistemic moments earlier time. Of course, it is possible
to formulate the corresponding variant of GSSC: Let E&V&W be a general time-
observation partition from tn to tn+m over [ti,tn+m]. Then,
(GSSC-) Cn+m(X)=Cn(X in v
mp/D&R) if E&V&W is optimal and sufficiently
inclusive for X,
where D is the sequential de-indexicalization of E under V, and R is the sequential re-
indexicalization of W for the m epistemic moments earlier time. It is provable that GSJC-
is equivalent to GSJC and that GSSC- is equivalent to GSSC. (See APPENDIX A.)
127127127127
In GSJC- and GSSC
-, epistemic moments that are too far past are not mentioned
as long as “What times were it in those moments?” is an irrelevant question to how
probable the target tensed proposition is now. Hence, we now have more practical
variants of GSJC and GSSC.
G. Application to the SB Problem
So far, I have developed a more general rule for de nunc updating than SSJC. This
generalization was initially motivated by the fact that SSJC does not apply correctly to
SB’s credal transition from m to m+. Hence, it will be interesting to see how well our new
rule, GSJC, will do with respect to the same credal transition.
For an easier discussion, I will first apply GSJC to a variant of the SB problem.
The lessons from the variant problem will help us to understand how to apply GSJC to
the original problem. Think about this version of the SB problem: SB problem 3. SB is
born on Sunday, knowing that it is Sunday. She also knows the following events will
happen during the next three days: Immediately after her birth, a group of evil
experimenters put her to sleep. Next, they toss a fair coin. Case 1: (H) The coin lands
heads. Then, they wake her up once on Monday. Case 2: (T) The coin lands tails. In this
case, the experimenters wake her up twice, the first time on Monday and the second time
on Tuesday. Between the two awakenings, they inject her with a drug that erases her
memory of the first awakening. In either case, SB is told that it is Monday one minute
after she wakes up on Monday. Here is the question: What is her credence in H when she
is told that it is Monday?
Let s be the moment of her birth on Sunday, m be the moment of her wakeup on
Monday, and m+ be the moment of being told that it is Monday. There are two ways in
128128128128
which we can answer the above question by using GSJC. First, we can use GSSC for her
credal transition from m to m+:
(25) Cm+(H)=Cm(H on Monday/(MON on Monday)&(MON at
prev0)&(SUN at prev1)).
Since H is a genuine proposition, [H on Monday] is equivalent to H. It is a tautology that
MON is true on Monday. Finally, when she wakes up on Monday, she remembers that it
was previously Sunday. Hence,
(26) Cm+(H)=Cm(H /MON at prev0)=Cm(H/MON)=
Cm(H&MON)/Cm(MON).
To find the last value, we use GSJC for her credal transition from s to m. At m, SB is sure
that she is observing W at that moment and that it was previously Sunday, but she does
not know whether it is Monday or Tuesday. Thus, her general time-observation partition
from s to m is W&(MON at prev0)&(SUN at prev1), W&(TUE at prev0)&(SUN at prev1).
Hence, the correct instance of GSJC is
(27) Cm(H&MON)=
Cs((H&MON) on Monday/(W on Monday)&(SUN at prev0))*
Cm(W&(MON at prev0)&(SUN at prev1))+
129129129129
Cs((H&MON) on Tuesday/(W on Tuesday)&(SUN at prev0))*
Cm(W&(TUE at prev0)&(SUN at prev1)).
Now observe the following facts: First, [(H&MON) on Monday] is clearly equivalent to
H. Second, on Sunday, SB knew that it was Sunday, and she fully expected to wake up
on Monday. Third, H&MON cannot be true on Tuesday. Thus, we can simplify (27) into
(28) Cm(H&MON)=Cs(H)*Cm(W&(MON at prev0)&(SUN at prev1)).
On Sunday, SB believed to the degree of ½ that the coin would land heads. On Monday,
she certainly knows that she is waking up with such and such a memory, and she
remembers that it was previously Sunday. Therefore,
(29) Cm(H&MON)=1/2Cm(MON at prev0)=1/2Cm(MON).
It follows from (26) and (29) that her credence at m+ in H is ½. (Note that this result
captures the intuition that I described in Section C.)
Second, GSSC provides the following instance for SB’s credal transition from s
to m+: When SB is told that it is Monday, she remembers that she previously experienced
waking up with the memory of Sunday as the last memory, and she learns that it is
Monday now. Hence, she is sure at m+ of (MON at prev0)&(W at prev1). Also, she is sure
at that moment that it is Monday then and that it was Monday previously. Thus, she is
130130130130
sure at m+ of (MON at prev0)&(MON at prev1). Finally, she remembers that it was
Sunday two epistemic moments ago. So, she is sure at m of (SUN at prev2). Therefore,
(30) Cm+(H)=Cs(H on Monday/(MON on Monday)&(W on Monday)&
(SUN at prev0)),
because (MON on Monday)&(W on Monday) is the sequential de-indexicalization of
(MON at prev0)&(W at prev1) under (MON at prev0)&(MON at prev1), and (SUN at prev0)
is the re-indexicalization of (SUN at prev2) for the two epistemic moments earlier time.
Of course, she fully expected on Sunday night that MON would be true on Monday and
that she would wake up on Monday. Also, she knew on that night that it was Sunday. As
a result,
(31) Cm+(H)=Cs(H on Monday)=Cs(H)=1/2.
Therefore, we arrive at the same conclusion whether we apply GSJC to SB’s credal
transition from s to m and then apply it to that from m to m+ step by step, or we apply it
to her credal transition from s to m+ all at once.
This is an intuitive result, and the answer to the given question resembles the
traditional Thirder view of the same question regarding the original SB problem. My next
question is whether we can apply GSJC to the credal transitions in the original SB
problem and acquire the same credence of hers at m+ in H.
131131131131
Here, we are faced with a difficulty. In the original SB problem, it was not
explicitly stated when SB made observations before the experiment began on Sunday
night. Since that information is crucial for using GSJC, we cannot apply that rule to the
original version of the SB problem.
What do we do? We can use GSJC- instead. Here is the rough idea: Perhaps,
there are many possibilities regarding when SB made observations before Sunday night.
However, we can safely assume that those possibilities are irrelevant to how the coin
lands on Monday. Under this assumption, we can apply GSJC- and GSSC
- without
worrying about the mentioned possibilities about what times it had been until Sunday
night.
To precisify this idea, let Rsoo∈O be the partition whose members describe
what days it had been at the epistemic moments before s, where Rso=(D
1o at prev1)&
(D2
o at prev2)&(D3
o at prev3)& … . Let Rmoo∈O be a similar partition such that for
each o∈O, Rmo=(D
1o at prev2)&(D
2o at prev3)&(D
3o at prev4)& … . I make these
assumptions:
(32) For any o∈O, Cs(H/W on Monday)=Cs(H/(W on Monday)&Rso)
and Cs(H/W on Tuesday)=Cs(H/(W on Tuesday)&Rso).
(33) For any o∈O, Cm(H/MON)=Cm(H/MON&Rmo).
54
54 I am assuming that Rs
o is well-specified de priori information with respect to <Cs, H, W on Monday>
and <Cs, H, W on Tuesday>. Similarly, Rmo is well-specified de priori information with respect to <Cm, H,
MON>.
132132132132
In words, from the point of view at s, what days it had been before now is irrelevant to H
conditional on [W on Monday] and, from the point of view at m, what times it had been
before the previous moment is irrelevant to H conditional on MON. These are highly
plausible assumptions. After all, both Rso and Rm
o are tensed propositions describing
what times it had been before Sunday night. Clearly, we have no reason to think that such
a matter is relevant in judging the probability of the coin’s landing heads.
Given these assumptions, I apply GSJC- to SB’s credal transition from s to m:
Consider partition W&MON&(SUN at prev1)&Wm
oo∈O∪W&TUE&(SUN at prev1)&
Wmoo∈O, where Wm
o=(D1
o at prev2)&(D2
o at prev3)&(D3
o at prev4)& … . By definition,
it is SB’s general time-observation partition from s to m.55
Also by definition,
W&MON&(SUN at prev1), W&TUE&(SUN at prev1) is a general time-observation
partition from s to m over [s,m]. Furthermore, the doubleton is sufficiently inclusive.56
Since it is the same partition I used for the credal transition from s to m in SB problem 3,
(27) is also an instance of GSJC- for the credal transition from s to m in the original. By
the same reasoning as in (27)-(29), Cm(H&MON)=1/2Cm(MON), i.e, (29) is also true in
the original SB problem.
55 From SB’s point of view at m, W describes her present observation, (MON at prev0) and (TUE at prev0)
describe the days that are possibly today, and (SUN at prev1)&Wm
o describe what times it had been until the
previous moment.
56
By (32), Cs(H/W on Monday)=Cs(H/(W on Monday)&Rso) and Cs(H/W on Tuesday)=Cs(H/(W on
Tuesday)&Rso) for any o∈O. Since CS(SUN at prev0)=1, Cs(H/(W on Monday)&(SUN at prev0))=Cs(H/(W
on Monday)&(SUN at prev0)&Rso) and Cs(H/(W on Tuesday)&(SUN at prev0))=Cs(H/(W on
Tuesday)&(SUN at prev0)&Rso) for any o∈O. Since (W on Monday) is the re-indexicalization of W under
MON and [(SUN at prev0)&Rso] is the sequential re-indexicalization of ((SUN at prev1)&W
mo) for each
o∈O, W&MON&(SUN at prev1), W&TUE&(SUN at prev1) is sufficiently inclusive, by definition. Done.
133133133133
Next, I apply GSSC- to SB’s credal transition from m to m+. Consider this
partition: MON&(MON at prev1)&(SUN at prev2)&Wm+
oo∈O, where Wm+o=(D
1o at
prev3)&(D2
o at prev4)&(D3
o at prev5)& … . By definition, it is SB’s general time-
observation partition from m to m+.57
Also, think about MON&(MON at prev1)& (SUN
at prev2). By definition, this singleton is her general time-observation partition from m
to m+ over [s,m+]. Moreover, the second partition is sufficiently inclusive.58
Therefore,
(25) is an instance of GSSC- for the credal transition from m to m+. By the same
reasoning as from (25) to (26) in SB problem 3, Cm+(H)= Cm(H&MON)/Cm(MON), i.e.,
(26) is also true in the original SB problem. From (26) and (29), it follows that
Cm+(H)=1/2.
Finally, I apply GSSC- to SB’s credal transition from s to m+. Consider this
partition: [(MON at prev0)&(W at prev1)] & [(MON at prev0)&(MON at prev1)] &
Wm+oo∈O. By definition, it is SB’s general time-observation partition from s to m+.
59
Plus, consider [(MON at prev0)&(W at prev1)] & [(MON at prev0)&(MON at prev1)].
By definition, this singleton is a general time-observation partition from s to m+ over
57 From B’s point of view at m+, MON describes her present observation and time, and (MON at prev1)&
(SUN at prev2)&Wm+
o describes what times it had been until the previous moment.
58
By (33), Cm(H/MON)=Cm(H/MON&Rmo). Since Cm(SUN at prev1)=1, Cm(H/(MON on Monday)&(MON
at prev0)&(SUN at prev1))=Cm(H/(MON on Monday)&(MON at prev0)&(SUN at prev1)&Rm
o). Since (MON
on Monday) is the de-indexicalization of MON under MON and (MON at prev0)&(SUN at prev1)&Rm
o is
the sequential re-indexicalization of (MON at prev1)&(SUN at prev2)&Wm+
o for the one epistemic moment
earlier time, MON&(MON at prev1)& (SUN at prev2) is sufficiently inclusive. 59
From B’s point of view at m+, [(MON at prev0)&(W at prev1)] describes what observations she has made
since the previous moment, [(MON at prev0)&(W at prev1)] describes what days it has been since the
previous moment, and Wm+o describes what days it has been since the two epistemic moments earlier time
(from which she is updating).
134134134134
[s,m+]. Also, the singleton is sufficiently inclusive.60
Hence, (30) is also an instance of
GSSC-. Consequently, Cm+(H)=1/2.
In summary, we can apply GSJC-, or equivalently GSJC, to SB’s credal
transitions from s to m and from m to m+ under some highly plausible assumptions, and
that application leads to coherent results that comply with the popular Thirder view. This
supports my suspicion that GSJC is the correct rule applicable to an agent’s credal
transition from when she did not know what time it was.
H. The Relation between GSJC and Other Rules
In this section, I discuss the relation between GSJC and other rules discussed in the
earlier chapters. First, I will review RSSJC*, a restricted version of SSJC discussed in the
last chapter. Second, I will present RSSJC, which is directly derivable from GSJC, and
compare it with RSSJC*. Finally, I will discuss how the restricted versions of various
rules in this dissertation are derived from GSJC.
First, remember this rule: Let Eo&Voo∈O be B’s sequential time-observation
partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k) and Vo=&1≤k≤m(V
ko at prevm-k).
Then, for any tensed proposition X,
60 By (32), Cs(H/W on Monday)=Cs(H/(W on Monday)&Rs
o) for any o∈O. Since CS((MON on
Monday)&(SUN at prev0))=1, Cs(H/(MON on Monday)&(W on Monday)&(SUN at prev0))=Cs(H/(MON on
Monday)&(W on Monday)&(SUN at prev0)&Rso). By definition, (MON on Monday)&(W on Monday) is
the sequential de-indexicalization of (MON at prev0)&(W at prev1) under (MON at prev0)&(W at prev1), and
(SUN at prev0) and ((SUN at prev0)&Rso) are the sequential re-indexicalizations for the two epistemic
moments earlier time. By definition, the above singleton is sufficiently inclusive. Done.
135135135135
(RSSJC*) Cn+m(X)= ∑o∈OCn(X in vm
o/Do)Cn+m(Eo&Vo) if Eo&Voo∈O is logically
optimal for X, and B was free from temporal ignorance at tn,
where Do is the sequential de-indexicalization of Eo under Vo for each o∈O. Here,
freedom from temporal ignorance at tm means that at tn B knew what time it was then. It
does not imply that at tn B knew anything about what time it was at the even earlier
moments (which we know to be tn-1, tn-2, etc.). Since this is a restrictive rule, it is
reasonable to expect that RGSJC* turns out to be a sub-principle of the new general rule,
GSJC, i.e., all the instances of RGSJC* turn out to be those of GSJC.
Unfortunately, the proviso of RGSJC* is neither strong enough to guarantee that
its instances are all derivable from GSJC nor weak enough to share all interesting
instances of GSJC. To see this point, second, look at this rule: Let Eo&Voo∈O be B’s
sequential time-observation partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k) and
Vo=&1≤k≤m(Vko at prevm-k). Note that Eo&Voo∈O is also B’s general time-observation
partition from tn to tn+m over [tn+1,tn+m]. Then, for any tensed proposition X,
(RSSJC) Cn+m(X)= ∑o∈OCn(X in vm
o/Do)Cn+m(Eo&Vo) if Eo&Voo∈O is logically
optimal and sufficiently inclusive for X,
where Do is the sequential de-indexicalization of Eo under Vo for each o∈O. Clearly, this
rule directly follows from GSJC-, which is equivalent to GSJC.
136136136136
To see the difference between RSSJC* and RSSJC, think about the logical
relation between freedom from temporal ignorance and sufficient inclusion. Clearly,
freedom from temporal ignorance is neither necessary nor sufficient for the given
partition’s being sufficiently inclusive. For Eo&Voo∈O is sufficiently inclusive for X
exactly when (#) Cn(X in vm
o/Do)=Cn(X in vm
o/Do&Ro) for any o∈O, where Do is the
sequential de-indexicalization of Eo under Vo and Ro is the sequential re-indexicalization
of Wo, the complement of Eo&Vo, for the m epistemic moments earlier time. To satisfy
(#), either (i) B must have completely known at tn what times it had been until then or (ii)
her credence distribution at tn must have been such that, for any o∈O, what times it had
been is irrelevant to whether X will true in vm
o conditional on what observations she will
make when. Even if B knew at tn what time it was then, it is not sufficient to satisfy (i),
and even if B did not know what time it was at tn, (#) still can be satisfied if (ii) holds.
Given these facts, I find RSSJC to be more attractive as a restricted version of
SSJC. Similar points hold for the other rules that I have discussed: Let E&V be B’s
general time-observation partition from tn to tn+m over [tn+1,tn+m]. Then, for any tensed
proposition X,
(RSSSC) Cn+m(X)=Cn(X in vm/D) if E&V is logically optimal and sufficiently
inclusive for X,
where D is the sequential de-indexicalization of E under V. Let Eo&Voo∈O be B’s time-
observation partition from tn to tn+1. It is also her general time-observation partition from
tn to tn+1 over [tn+1,tn+1]. Then,
137137137137
(RSJC) Cn+1(X)=∑o∈OCn(X in vo/Eo in vo)Cn+m(Eo&Vo) if Eo&Voo∈O is logically
optimal and sufficiently inclusive for X.
Let E&V be B’s time-observation partition from tn to tn+1. Then,
(RSSC) Cn+1(X)=Cn(X in v/E in v) if E&V is logically optimal and sufficiently
inclusive for X.
These principles provide convenient shortcuts for the cases where the direct application
of GSJC is awkward.
When restricted in this way, I believe that SSJC, SSSC, SJC, and SSC become
more safe and attractive rules. As such, GSJC subsumes the intuitions behind the
attractive instances of these rules.
I. Conclusion
At this point, it will be useful to recapitulate the discussions I have presented thus far in
this dissertation. In Chapter II, I developed an updating rule, SJC, which provides an
elegant solution for the SB problem. In Chapter III, I generalized that rule for the credal
transitions between epistemic moments that are not necessarily contiguous. Unfortunately,
the resulting rule, SSJC, yielded different results depending upon whether we apply it to
SB’s credal transitions step-by-step (from s to m and then from m to m+) or all at once
(from s to m+). In this chapter, I presented a rule free from such inconsistency, at least
regarding the SB problem. If I had suggested modifying SSJC to GSJC only to avoid the
138138138138
inconsistent results, I would not be able to avoid the charge of adhocery; however, I have
avoided this charge by providing independent reasons for such modification.
139139139139
CHAPTER V
SATISFACTION OF DESIDERATA
A. Introduction
In the previous chapters, I discussed a series of updating rules. In each chapter, I
suggested a rule for updating one’s de nunc credences (or the degrees of tensed beliefs).
In each case, however, a problem was discovered. Facing each problem, I responded by
suggesting an enhanced rule, immune to the newly discovered problem. In addition, I
provided an independent reason to favor the enhanced rule, on the basis of Gaifman’s
Expert Principle.
In chapter IV, I presented GSJC as the final product of this process. It supports
some instances of SJC and SSJC, which meet a few special conditions. If we restrict SJC
and SSJC with those conditions as provisos, as I believe that we should, we may consider
them to be subordinate rules of GSJC. As such, GSJC is the most general rule discussed
so far.
This raises a question: Is GSJC the most general rule for de nunc updating, full-
stop? In other words, is GSJC a rule such that it is always rational to update one’s de
nunc credences in accordance with it (hereafter: the Final Rule)?61
Recalling my trial and error in the earlier chapters, I find this question to be hard
to answer with confidence. For how can we rule out the possibility that GSJC also suffers
61 Of course, other conditions will have to be satisfied. For example, I will assume that the agent has a
perfect memory about her own past opinions. In this chapter, I will assume that the given agent satisfies
such basic conditions.
140140140140
from a now unknown but serious problem and that there exists a better rule for updating,
which is immune to that problem and defensible for other reasons? Of course, we cannot.
Nevertheless, GSJC might be the Final Rule. Think about the following
properties, which any acceptable rule for updating will satisfy. First, the Final Rule will
produce only synchronically coherent credence functions that diachronically cohere with
the earlier ones: Suppose that the agent has always updated her de nunc credences in
accordance with the Final Rule. Then, her resulting present credence function will satisfy
the standard axioms of probability, and it will be rationally related to any past credence
distribution of the same agent’s. Second, the Final Rule will be observationally
exhaustive: If the agent has updated her de nunc credences in accordance with the Final
Rule, the present credal judgments resulting from such updates will incorporate the
totality of what she has observed until now.
In this chapter, I argue that GSJC satisfies these two requirements. Hence, GSJC
is not disqualified as a candidate for the Final Rule; at least, it is not disqualified due to
any failure to satisfy those requirements. This may not seem significant, but remember
that I already offered a general argument for GSJC. Plus, it generates plausible instances
for various types of credal transitions. In my opinion, when combined, these facts form
good evidence that GSJC is very close to the Final Rule, if not identical.
B. Background
Before I begin the main discussion, I will clarify several notions and assumptions that I
will use and make in this chapter.
First, I clarify the notions of synchronic and diachronic coherences: To some
extent, these words are self-explanatory: An agent’s credence function at tn is
141141141141
synchronically coherent iff the elements of that function cohere with one another, and it
diachronically coheres with the earlier credence functions iff it is rationally related to the
agent’s credence functions at tn-1, tn-2, and so on (up to t0). However, the real challenge is
to provide substantial criteria for such coherence, within the same credence function and
between different credence functions at different times.
If we put aside a few thorny matters (such as the status of Countable Additivity
as an axiom), it is relatively easy to find the criterion for synchronic coherence: An
agent’s credence function Cn at tn is synchronically coherent iff Cn satisfies Non-
negativity, Normality, and Addivity.62
Regarding de dicto credences, a comparable
criterion will be SC or JC: An agent’s credence function Cn diachronically coheres with
her earlier credence functions iff Cn is related by SC/JC to each of Cn-1, …, C0.
Concerning de nunc credences, a comparable criterion will be, hopefully, GSJC: An
agent’s credence function Cn diachronically coheres with the earlier credence functions
iff Cn is related by GSJC to each of Cn-1, …, C0.
Second, I discuss how to apply the notion of transitivity to a rule for updating.
We all know what transitivity is: For any binary relation R and its field S, R is transitive
iff for any x,y,z∈S, if x is related by R to y and y is related by R to z, then x is related by R
to z. However, it is not so obvious that a rule for updating captures a binary relation.
Typically, such a rule relates more than two things. A rule for updating describes the
relation between two credence functions, but its relata include other entities as well. For
example, SC/JC seems to describe the relation among (i) the agent, (ii) her observations
at the given moment, and (iii) her old and new credence functions. However, we can
62 For simplicity, I shall use “Additivity” to refer to Finite Additivity in this chapter. However, everything
discussed here would apply in the same way if I were to use that word to refer to Countable Additivity.
142142142142
focus upon a fixed agent and her history of experiences. With those other elements fixed,
we can regard SC/JC as (capturing) a binary relation between two credence functions.
(Compare: The ancestor-descendant relation appears to have a family as one of the
related entities, but we can treat it as a binary relation by considering only a fixed family
line.) By using a similar method, we can define transitivity for updating principles: Focus
upon a fixed agent B (having such and such an epistemic history H). Then, an updating
rule R is transitive iff for any credence functions Cn, Cn+m, Cn+m+l of B’s, if Cn+m is related
by R to Cn, and Cn+m+l is related by R to Cn+m, then Cn+m+l is also related by R to Cn.
Third, I discuss the notion of “epistemic kernel rules for updating”: As I
mentioned above, a rule for updating describes a relation connecting (at least) an agent’s
credence functions at different moments. Usually, neither of those credence functions is
assumed to play a special role in that relation. For example, think about
“Cn+1(X)=Cn(X/E)” as an instance of SC. In the credal transition described here, we can
say that Cn is the source and Cn+1 is the result but, most likely, Cn was the result of the
given agent’s previous credal transition, and Cn+1 will be the source for her next credal
transition. In this sense, Cn and Cn+1 are not so different in their roles. Let’s call this
category of rule “transitional rules for updating.”
Interestingly, Meacham (2008; forthcoming) suggests a different type of rule for
updating, which he calls “epistemic kernel rules for updating.” According to him, an
epistemic kernel rule for updating relates an agent’s ordinary credence function to a
special credence function, which he calls “the kernel.” When compared with ordinary
credence functions, a kernel is supposed to play a special role. For example, here is an
epistemic kernel version of SC: Cn(X)=C0(X /TE) for any proposition X, where C0 is the
143143143143
agent’s kernel and TE is the totality of her observations until tn. Note that every instance
of this rule includes C0 as the source and never includes it as the result. In this sense, C0
plays a different role from that of any other credence function of the same agent’s.
Exactly what is this so-called kernel? It is difficult to say, because Meacham
does not provide a very clear definition for that notion. Here is my best shot: Your kernel
is a credence function that you would have if you were stripped of all the data from your
past and present observations. For this reason, if you update your credences in
accordance with an epistemic kernel rule, it will transform your kernel into a new
credence function, incorporating the totality of all your observations whatsoever until
now, not just the totality of your observations after some earlier epistemic moment.
While this is an interesting idea, it comes at a cost. First, one may complain that
an agent’s observations might be so crucial for making any credal judgment that if she
were stripped of them, she would not have any remaining credal opinions. Hence, there is
no guarantee that every agent has a kernel in the above sense. Second, “if she were
stripped of all her observational data” seems to be figurative language. How would we
express the above idea more literally? I am not sure, and I suspect that many people will
feel the same way.
To avoid these problems, I adopt this approach: I will consider a set of special
agents, those who had their first credal opinions. Hence, each of them is guaranteed to
have an initial credence function C0, and C0 was literally formed without the help of any
observational data. By considering only those having these features, we can avoid the two
problems mentioned in the previous paragraph.
144144144144
Of course, this approach has its own cost: If we adopt this approach, our
discussion about epistemic kernel rules will be restricted to the credence functions had by
these special agents. In my opinion, this is a quite acceptable cost. For first, almost every
possible agent must have had an initial credence function at some point in her life, except
those rare agents who have lived for an eternal time with no beginning and, for those not
in this elite group, their initial credences are likely to have been based upon very few
observations. Moreover, second, what we can learn from this potentially narrow range of
agents may include important lessons applicable to a broader range of agents, especially
about a special topic such as the rational rule for updating de nunc credences.
In this section, I have clarified the notions and assumptions that I will use
moving forward. Having done this, I am now ready for the main discussions.
C. Strategy
Earlier, I suggested that if an agent has updated her credences in accordance with the
Final Rule, first, her resulting credence function will be synchronically coherent and it
will diachronically cohere with her earlier credence functions, and second, as a form of
judgment, that credence function will incorporate the totality of her observations until the
present moment.
Why do I think that the Final Rule will satisfy these requirements? It is clear why
a rational rule for updating will produce only synchronically coherent credence functions:
For any single credence function of yours, you want its elements to be rationally related.
The standard axioms of probability are meant to capture exactly this relation (in the
domain of subjective probabilities).
145145145145
It is also clear why a rational rule for updating will produce only a credence
function that diachronically coheres with any earlier credence function: For any two
credence functions of yours, you want one of them to be rationally related to the other.
For even if your credence function is internally coherent at any fixed time, you will be
regarded as an unacceptably whimsical agent unless your credence function at each
moment is rationally related to those at the earlier moments. The standard rules for de
dicto updating, strict and Jeffrey conditionalizations, are meant to capture this relation. In
this dissertation, I am trying to find a similar rule for de nunc updating.
How about the requirement of observational exhaustiveness? Many mainstream
epistemologists will agree that an agent’s judgment needs to be based upon what she has
observed. The following thesis, especially, is popular in the literature: An agent’s belief is
justified iff that belief is supported by her evidence. (Connee and Feldman (2004) offer a
similar thesis). I also find it to be an intuitive claim.
However, we should be careful. First, in the literature, “evidence” has been used
frequently with the connotation that its referent is easily accessible to the given agent
(hereafter: accessibility connotation). Perhaps, we can formulate a similar thesis without
being committed to this connotation or sacrificing the intuitive appeal of the above thesis.
For this formulation, I suggest using “observation” to refer to what plays a similar
justificatory role to that of evidence but does not have the accessibility connotation.
Second, even if an observation E supports a hypothesis H when considered in
separation, the rest of one’s observations may include other information that undercuts or
rebuts E’s support for H (Pollock & Cruz, 1986). Hence, if we judge whether an agent’s
belief is justified by considering only fragments of her observations, it may lead to a
146146146146
disastrous result. Therefore, we should make such a judgment by considering the totality
of the given agent’s observations.
These considerations lead to the following refinement of the above thesis: An
agent’s belief is justified iff that belief is justified by the totality of her observations. Can
we formulate an analogous thesis for probabilism? Yes: An agent’s degree of belief is
justified iff that degree of belief is supported by the totality of her observations. This
thesis is plausible for the same reason as its counterpart in the mainstream epistemology:
Intuitively, a rational agent’s credal judgment needs to be based upon her observations,
and if she makes such a judgment on the basis of only a proper subset of her observations,
the rest of her observations may include what would have resulted in a different credal
judgment if considered in making that judgment. In addition, the above thesis is
compatible with the updating model based on JC, in which the agent is not assumed to
have infallible access to her own observations.63
For these reasons, I believe that any ideal general rule for de nunc updating will
satisfy the two mentioned requirements. Of course, even if a certain rule for de nunc
updating, say, R, satisfies those requirements, it is still possible that R fails to satisfy
some other crucial requirements, of which I am not aware yet. However, R’s satisfaction
of the discussed requirements will certainly provide some reason to suspect that R is the
general rule for de nunc updating.
In the rest of this chapter, I shall proceed in the following order: In Section D, I
will present GSJC and GSR in yet new forms. As formulated in these new forms, they
63 Another merit is its compatibility with justificatory externalism, but this does not look like a significant
merit for the standard probabilism, which already assumes a psychological (read: “inside the skull”)
account of beliefs. However, see Williamson (2002) for an externalist version of probabilism.
147147147147
will be called “GSJC+” and “GSR
+.” In Section E, I will prove that GSR
+ captures a
transitive binary relation between credence functions, a crucial lemma for the rest of the
chapter. In Section F, I will show that GSJC+ satisfies the first requirement: If you have
always updated your credences in accordance with GSJC+, your resulting credence
function will be synchronically coherent, and it will diachronically cohere with your
earlier credence functions. In Section G, I will argue that GSJC+ satisfies the second
requirement: Under the same assumption, we can show that your credence functions
resulting from such updates will incorporate the totality of your past observations. In
Section H, I will point out a difference between GSJC, the earlier formulation of my
general updating rule, and GSJC+, the new formulation of that rule presented in this
chapter. I will suggest a way to fill this gap.
D. GSJC+ and GSR
+
In this section, I present GSJC and GSR again, but this time, in a more technically
rigorous way. After that, I discuss their logical relations.
In presenting GSJC and GSR in new forms, I put two restrictions on them: First,
I put a restriction on how the domain of credence functions is constructed. In the earlier
chapters, I didn’t impose any restriction on it except that it consists of tensed propositions.
However, it is usual to assume algebra made out of propositions, events, or sets as the
domain of credence functions. Since we are dealing with tensed propositions, let Γ be the
class of all tensed propositions, and let Ω be an algebra made out of S, i.e., Ω is a subset
of the power set of Γ closed under conjunction and complementation. Additionally, we
assume that Ω is also closed under [… at τ] and [… in ν] operations, i.e., if X is a
member of Ω, then so are [X at τ] and [X in ν], where “τ” is any term referring to a
148148148148
moment, and “ν” is any term referring to an interval. Next, let ∆ be the set of an agent B’s
credence functions. We assume that for any C∈∆, the domain of C is Ω.
Second, I put a restriction on the time intervals and time-specifying tensed
propositions, which will appear in my new formulations of GSJC and GSR. Consider
partitions Φ and Ψ⊆Ω, which meet the following conditions: (i) Φ consists of temporal
intervals, and Ψ consists of tensed propositions specifying what time it is. (ii) For any
v∈Φ, Ψ includes V or the tensed proposition that it is v now. Conversely, for any V∈Ψ,
Φ includes a minimal interval throughout which V is true. (iii) For any X∈Ω and v∈Φ,
the truth-value of X is invariant within v. (iv) Let X and Y be any genuine propositions
that belong to Ω. Suppose that r=C(X/Y&R), where C∈∆ and R=(W1 at prev0)&(W2 at
prev1)&…&(Wm at prevn) for some <W1
, W2,…, W
n>∈Ψn
. Hence, R is a temporal
description or a tensed proposition thoroughly describing what times it has been until
now. In this case, even if we replace R with a better temporal description R*, still
C(X/Y&R*)=r.64
(To see what this means practically, suppose that r=Cn(X in vm/D&R), where
Cn∈∆, X∈Ω, D=(E1 in v
1)&…&(E
m in v
m) for some <E
1…,E
m>∈Ωm
and <v1…,v
m>∈ Φm
,
and R=(W1 at prev 0)&…&(W
m at prevn+1) for some <W
1,…,W
n+1>∈Ψn+1
. By (iii), the
truth-value of X is invariant within vm, that of E
1 is invariant within v
1, …, and that of E
m
is invariant within vm. By (iv), Cn(X in v
m/D&R) is conditioned upon a well-specified
64 Remember the definition of better de priori information in the last chapter: Let R be (W
1 at prev
0)&…&(Wm at prevn) and R* be (W*
1 at prev 0)&…&(W*
m at prevn). Let w
1,...,w
n be intervals associated
with W1,...,W
n; similarly for w*
1,...,w*
n. Then, R* is a better temporal description than R iff (i) for every
k∈1,...,n,wk⊇w∗k
and (ii) for some k∈1,...,n, wk⊃w∗k
.
149149149149
temporal description. Clearly, this allows us to do away with the provisos of optimality in
the earlier formulations of GSJC and GSR.)
Whenever ∆, Ω, Φ, and Ψ satisfy these conditions, I will say that <∆,Ω,Φ,Ψ> is
a model for B’s (de nunc) credences. Given a model <∆,Ω,Φ,Ψ> for B’s credences, we
are ready to formulate the wanted principles: As before, we let Eo&Vo&Woo∈O be an
agent B’s general time-observation partition from tn to tn+m, where Eo=&1≤k≤m(Eko at
prevm-k), Vo=&1≤k≤m(Vko at prevm-k), and Wo=&1≤k≤n+1(W
ko at prevm+k-1). The difference is
that this time, we assume that, for each o∈O, <E1
o,…Em
o>∈Ωm, <V
1o,…,V
mo>∈Ψm
, and
<W1
o,…,Wn
o>∈Ψn. (Hereafter, I will say that Eo&Vo&Woo∈O is constructed from Ω and
Ψ whenever it meets these conditions.) Let Cn, Cn+m∈∆. Consider any X∈Ω. Then, I
make these claims:
(GSJC+) Cn+m(X)=Σo∈OCn(X in v
mo/Do&Ro)Cn+m(Eo&Vo&Wo).
(GSR+) Cn+m(X/ Eo&Vo&Wo)=Cn(X in v
mo/Do&Ro) for each o∈O.
Here, Do is the sequential de-indexicalization of Eo under Vo, and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time (i.e., Do= &1≤k≤m(Eko in
vko) and Ro=&1≤k≤n+1(W
ko at prevk-1)).
150150150150
Next, I discuss the logical relation between GSJC+ and GSR
+: Let Cn and Cn+m
be B’s credence functions at tn and tn+m. (Hence, Cn, Cn+m∈∆.) Then, we can prove these
two facts:
(1) If Cn+m is synchronically coherent and related by GSR+ to Cn, then Cn+m is
also related by GSJC+ to Cn.
(2) If Cn is synchronically coherent and Cn+m is related by GSJC+ to Cn, then
Cn+m is also related by GSR+ to Cn.
(See APPENDIX B for the proofs.) It is true that these simply mean that GSJC+ is
equivalent to GSR+ because a rational agent’s credence function always will be
synchronically coherent. But it is one of my goals in this chapter to show that if you
always update in accordance with GSJC+, all your credence functions will be
synchronically coherent. Hence, I do not want to assume the synchronic coherence of B’s
credence functions from the outset.
Still, this means that those principles will turn out to be equivalent, if we
independently prove the synchronical coherence of all B’s credence functions. Assuming
such an independent proof, it will be possible to use GSR+ as a proxy for GSJC
+: We can
show that GSJC+ satisfies a wanted requirement by showing that GSR
+ satisfies it.
E. The Transitivity of GSR+
In this section, I argue that GSR+ captures a transitive binary relation between one’s
credence functions at various moments. For an easier discussion, I will first prove an
151151151151
analogous claim for Rigidity, and then provide a proof for the transitivity of GSR+, by
modifying the first proof.
First, I prove the transitivity of (the de dicto version of) Rigidity: If Rigidity
holds between Cn and Cn+m and between Cn+m and Cn+m+l, then it also holds between Cn
and Cn+m+l. To prove this claim, we first suppose these conditions: Let Eii∈I* be a
partition such that each Ei describes a possible course of B’s observations during
[tn+1,tn+m]. Consider Eii∈I such that I⊆I*, Cn+m(Ei)>0 and Σi∈I Cn+m(Ei)=1. In such a
case, we will call Eii∈I “(B’s) observation partition from tn to tn+m.” Let Fjj∈J be also
B’s observation partition from tn+m to tn+m+l in a similar sense. We suppose that
(3) Cn+m(X/Ei)=Cn(X/Ei) for any proposition X and any i∈I, and
(4) Cn+m+l(X/Fj)=Cn+m(X/Fj) for any proposition X and any j∈J
and show that
(5) Cn+m+l(X/Gk)=Cn (X/Gk) for any proposition X and for any k∈K,
where Gkk∈K is B’s observation partition from tn to tn+m+l.
To show this, we need to know how Gkk∈K is related to Eii∈I and Fjj∈J. I
claim that each Gk should be Ei&Fj for some <i,j>∈Ι×J. For remember that each Ei
describes a possible course of observations during [tn+1,tn+m], and each Fk describes a
possible course of observations during [tn+m+1,tn+m+l]. Since each Gk represents a possible
152152152152
course of observations during [tn+ 1,tn+m+l], Gk=Ei&Fj for some <i,j>∈Ι×J. For
convenience, we define Ek=Ei and Fk=Fj when Gk=Ei&Fj. So it suffices to show
(6) Cn+m+l(X/Ek&Fk)=Cn(X/Ek&Fk) for any proposition X and any k∈K.
This is easy to prove: Let X be any genuine proposition. Then, we can derive the
following facts:
(7) Cn(X/Ek&Fk)=Cn(X&Fk/Ek)/Cn(Fk/Ek),
(8) Cn+m(X&Fk/Ek)/Cn+m(Fk/Ek)=Cn+m(X&Ek/Fk)/Cn+m(Ek /Fk), and
(9) Cn+m+l(X&Ek/Fk)/Cn+m+l(Ek/Fk)=Cn+m+l(X/Ek&Fk).
By (3), Cn(X&Fk/Ek)=Cn(X&Fj/Ei)=Cn+m(X&Fj/Ei)=Cn+m(X&Fk/Ek) and Cn(Fk/Ek)=
Cn(Fj/Ei)=Cn+m(Fj/Ei)=Cn+m(Fk/Ek). By (4), Cn+m(X&Ek/Fk)=Cn+m(X&Ei/Fj)=Cn+m+l(X&
Ei/Fj)=Cn+m+l(X&Ek/Fk) and Cn+m(Ek/Fk)=Cn+m(Ei/Fj)=Cn+m+l(Ei /Fj)=Cn+m+l(Ek/Fk). Done.
Second, I prove GSR+’s transitivity. My proof of this fact will be structurally
similar to that of the transitivity of Rigidity, although more complex: Let
Eo&Vo&Woo∈O be B’s general time-observation partition from tn to tn+m, where
Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and Wo=&1≤k≤n+1 (W
ko at prevm+k-
1). Also, let Fp&Vp&Wpp∈P be B’s general time-observation partition from tn+m to tn+m+l,
where Fp=&1≤k≤l(Fkp at prevl-k), Vp=&1≤k≤l (V
kp at prevl-k), and Wp=&1≤k≤n+m+1(W
kp at
153153153153
prevl+k-1). Both are constructed from Ω and Ψ. Then, we suppose that GSR+ holds
between Cn and Cn+m and between Cn+m and Cn+m+l. In other words,
(10) Cn+m(X/Eo&Vo&Wo)=Cn (X in vm
o/Do&Ro) for any tensed
proposition X and o∈O, and
(11) Cn+m+l(X/Fp&Vp&Wp)=Cn+m (X in vlp/Dp&Rp) for any tensed
proposition X and p∈P,
where Do is the sequential de-indexicalization of Eo under Vo, and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time, and Dp is the sequential
de-indexicalization of Fp under Vp, and Rp is the sequential re-indexicalization of Wp for
the l epistemic moments earlier time. From these suppositions, we prove that GSR+ holds
between Cn and Cn+m+l. Let Gq&Vq&Wqq∈Q be B’s general time-observation partition
from tn to tn+m+l, where Gq=&1≤k≤m+l(Gkq at prevm+l-k), Vq=&1≤k≤m+l(V
kq at prevm+l-k), and
Wq=&1≤k≤n+1(Wkq at prevm+l+k-1). We want to show that
(12) Cn+m+l(X/Gq&Vq&Wq)=Cn(X in vm+l
q/Dq&Rq) for any tensed
proposition X and q∈Q,
where Dq is the sequential de-indexicalization of Gq under Vq, and Rq is the sequential re-
indexicalization of Wq for the m+l epistemic moments earlier time.
154154154154
As before, the key to the proof is in finding the correct relation of
Gq&Vq&Wqq∈Q to Eo&Vo&Woo∈O and Fp&Vp&Wpp∈P. Here is that relation: For
any q∈Q, there exists <o,p>∈O×P such that
(13) Gq=&1≤k≤m(Eko at prevm+l-k)&&1≤k≤l(F
kp at prevl-k);
(14) Vq=&1≤k≤m(Vko at prevm+l-k)&&1≤k≤l(V
kp at prevl-k);
(15) Wq=&1≤k≤n+1(Wko at prevm+l+k-1) and Rq=&1≤k≤n+1(W
ko at prevk-1)=Ro;
(16) Wp=&1≤k≤m(Vko at prevm+l-k)&&1≤k≤n+1(W
ko at prevm+l+k-1) and
Rp=&1≤k≤m(Vko at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1)=Vo&Wo.
(See APPENDIX C for a proof.)
Having established these facts, I introduce the following definitions: For any
q∈Q, for the <o,p>∈O×P such that (13)-(16) hold between o, p, and q,
(17) Eq=df&1≤k≤m(Eko at prevm+l-k) and Fq=dfFp=&1≤k≤l(F
kp at prevl-k);
(18) VEq=df&1≤k≤m(Vko at prevm+l-k) and VFq= dfVp=&1≤k≤l(V
kp at prevl-k);
and
(19) DEq=dfDo=&1≤k≤m(Eko in v
ko) and DFq=dfDp=&1≤k≤l(F
kp in v
kp),
155155155155
It follows from (13)-(19) that, for any q∈Q,
(20) Gq=Eq&Fq, Vq= VEq&VFq, and Dq= DEq&DFq; and
(21) Wp=VEq&Wq.
Consider any X∈Ω. By (19), it suffices to show that
(22) Cn+m+l(X/Eq&Fq&VEq&VFq&Wq)=Cn(X in vm+l
q/DEq&DFq&Rq).
To show it, first note that
(23) Eo&Vo is equivalent to Do&Vo, and DEq&VEq is equivalent to
Eq&VEq.65
From the above definitions and facts, it follows that for any X∈Ω,
65 In showing the first equivalence, we are trying to show the equivalence between &1≤k≤m(E
ko at prevm+l-
k)&&1≤k≤m(Vko at prevm+l-k) and &1≤k≤m(E
ko in V
ko)&&1≤k≤m(V
ko at prevm+l-k). Clearly, it suffices to show that,
for any number k∈1,...,m, (Eko at prevm+l-k)&(V
ko at prevm+l-k) is equivalent to (E
ko in V
ko)&(V
ko at prevm+l-
k). To show this, consider any k∈1,...,m. (=>) Assume (Eko at prevm+l-k)&(V
ko at prevm+l-k) and show (E
ko
in Vko)&(V
ko at prevm+l-k). By supposition, it was v
ko at the m+l-k epistemic moments earlier time, and E
ko
was true then. By the construction of Ω and Φ, the truth-value of Eko is invariant within v
ko. Hence, E
ko is
true at any t∈vko; by definition, (E
ko in V
ko) is true. Since the supposition provides the other conjunct, done.
(<=) Assume (Eko in v
ko)&(V
ko at prevm+l-k) and show (E
ko at prevm+l-k)&(V
ko at prevm+l-k). By assumption,
Eko is true at any moment in v
ko and it was v
ko at the m+l-k epistemic moments earlier time. Clearly, it
follows that Eko was true at the m+l-k epistemic moments earlier time; by definition, (E
ko at prevm+l-k) is true.
Done. The second equivalence can be shown in a similar way.
156156156156
(24) Cn(X in vm+l
q/DEq&DFq&Rq)=
Cn((X in vm+l
q)&Dp/Do&Ro)/Cn(Dp/Do&Ro),66
(25) Cn+m((X in vm+l
q)&Dp/Eo&Vo&Wo)/Cn+m(Dp/Eo&Vo&Wo)=
Cn+m((X in vm+l
q)&Do/Dp&Rp)/Cn+m(Do/Dp&Rp),67
and
(26) Cn+m+l((X in vm+l
q)&Do/Fp&Vp&Wp)/Cn+m(Do/Fp&Vp&Wp)=
Cn+m+l(X/Eq&Fq&VEq&VFq&Wq).68
Thus, it suffices to show that
(27) Cn+m((X in vm+l
q)&Dp/Eo&Vo&Wo)=Cn((X in vm+l
q)&Dp/Do&Ro),
(28) Cn+m(Dp/Eo&Vo&Wo)=Cn(Dp/Do&Ro),
(29) Cn+m+l((X in vm+l
q)&Do/Fp&Vp&Wp)=Cn+m((X in vm+l
q)&Do/Dp&
Rp), and
(30) Cn+m+l(Do/Fp&Vp&Wp)=Cn+m(Do/Dp&Rp).
These facts follow from (10) and (11).69
Therefore, GSR+ is transitive.
66 For brevity, abbreviate “X in v
m+lq” into “X*.” Then, Cn(X*/DEq&DFq&Rq)=(by (15) and (18))
Cn(X*/Do&Dp&Ro)=Cn(Z1 &Do&Dp&Ro)/Cn(Do&Dp&Ro)=Cn(Z1&Dp/Do&Ro)/Cn(Dp/Do&Ro). Done.
67
Again, abbreviate “X in vm+l
q” into “X*.” Then, Cn+m(X*&Dp/Eo&Vo&Wo)/Cn+m(Dp/Eo&Vo&Wo)=(by
(22)) Cn+m(X*&Dp/Do&Vo&Wo)/Cn+m(Dp/Do&Vo&Wo)=(by (16)) Cn+m(X*&Dp/Do&Rp)/Cn+m(Dp/Do&Rp)=
Cn+m(X*&Dp&Do&Rp)/Cn+m(Dp& Do&Rp)=Cn+m(X*&Do/Dp&Rp)/Cn+m(Do/Dp&Rp). Done. 68
Again, abbreviate “X in vm+l
q” into “X*.” Then, Cn+m+l(X*&Do/Fp&Vp&Wp)/Cn+m(Do/Fp&Vp&Wp)=
Cn+m+l(X*&Do&Fp&Vp &Wp)/Cn+m(Do&Fp&Vp&Wp)=Cn+m+l(X*/Do&Fp&Vp&Wp)=(by (16)-(18) and (20))
Cn+m+l(X*/DEq&Fq&VEq&VFq&Wq)=(by (22)) Cn+m+l(X*/Eq&Fq&VEq&VFq&Wq). Done. 69
Abbreviate “(X in vm+l
q)&Dp” into “X1”and “(X in vm+l
q)&Do” into “X2.” It follows from (8) and (9) that
Cn+m(X1/Eo&Vo&Wo)=Cn(X1 in vm
q/Do&Ro), Cn+m(Dp/Eo&Vo&Wo )=Cn(Dp in vm
q/Do&Ro), Cn+m+l(X2/Fp&Vp
157157157157
This is an important result. Using it as a lemma, I will argue in the next two
sections that GSR+, when regarded as a rule for de nunc updating, satisfies the two
aforementioned requirements that must be met by any candidate for the Final Rule.
F. Synchronic and Diachronic Coherence
In this section, I argue that GSJC+ satisfies the first requirement for the Final Rule. In
other words, if one’s credences have been updated in accordance with GSJC+, the
resulting credence function is, first, synchronically coherent in itself, and second,
diachronically coherent with any earlier credence function.
For my argument, I assume of an agent B that (i) she always updates her
credences in accordance with GSJC+. Under this assumption, I will show the synchronic
and diachronic coherence of the resulting credence function of B’s. To show this, I
depend upon other assumptions as well: (ii) Her initial credence function was
synchronically coherent. (iii) The credence distribution over her general time-observation
partition is always synchronically coherent. Admittedly, neither is trivial, but both are
arguably fairly weak assumptions.70
(Certainly weaker than the results derived from the
assumptions.)
First, I show that B’s credence function is synchronically coherent at any
epistemic moment. Clearly, it suffices to prove that the synchronic coherence of B’s
&Wp)= Cn+m(X2 in vlp/Dp&Rp), and Cn+m(Do/Fp&Vp& Wp)=Cn+m(Do in v
lp/Dp&Rp). Since X1, Dp, X2, and
Do are all genuine propositions, “in vm
q” or “in vlq” is redundant. Done.
70 I find it especially difficult to explain why one’s credence distribution over the general time-observation
partition ought to be synchronically coherent. However, note that it is equally hard to explain why one’s
credence distribution over the observation partition for JC should be synchronically coherent.
158158158158
credence functions is preserved from tn to tn+1. To prove this, assume that Cn∈∆ is
synchronically coherent, i.e., for any X,Y∈Ω,
(NNn) Cn(X)≥0,
(NORMn) Cn(X)=1 if X is tautological, and
(ADDn) Cn(X ∨Y)=Cn(X)+Cn(Y) if ~(X&Y) is tautological.
From this assumption, we prove that Cn+1∈∆ is synchronically coherent, i.e., for any
X,Y∈Ω,
(NNn+1) Cn+1(X)≥0,
(NORMn+1) Cn+1(X)=1 if X is tautological, and
(ADDn+1) Cn+1(X ∨Y)=Cn+1(X)+Cn+1(Y) if ~(X&Y) is tautological.
To prove these facts, we will use the following theorems derivable from NNn, NORMn,
and ADDn: For any X,Y,Z∈Ω
(CNNn) Cn(X/Y)≥0 if defined,
(CNORMn) Cn(X/Y)=1 if defined and X is tautological, and
(CADDn) Cn(X∨Y/Z)=Cn(X/Z)+Cn(Y/Z)
if defined and ~(X&Y) is tautological.
159159159159
First, I show NNn+1: Let X be any tensed proposition, and let Eo&Vo&Woo∈O be B’s
general time-observation partition from tn to tn+1 (constructed from Ω and Ψ). By (i),
(31) Cn+1(X)=∑o∈OCn(X in v1o/Do&Ro)Cn+1(Eo&Vo&Wo),
where Do is the sequential de-indexicalization of Eo under Vo, and Ro is the sequential re-
indexicalization of Wo for the one epistemic moment earlier time. By (iii) and CNNn, all
terms on the right-hand side are non-negative. Hence, NNn+1 is true. Second, I show
NORMn+1: Consider any X that is tautological. Then, (X in v1
o) is also tautological; for, a
tautology is logically true at every moment in any temporal interval. By CNORMn, Cn(X
in v1o/Do&Ro)=1 for any o∈O. By the definition of the general time-observation partition,
∑o∈OCn+1(Eo&Vo&Wo)=1. Hence,
(32) Cn+1(X)=∑o∈OCn(X in v1
o/Do&Ro)Cn+1(Eo&Vo&Wo)=1.
Third, I show ADDn+1: Let X and Y be any members of Ω. Suppose that ~(X&Y) is
tautological. By (i),
(33) Cn+1(X)=∑o∈OCn(X in v1
o/Do&Ro)Cn+1(Eo&Vo&Wo),
(34) Cn+1(Y)=∑o∈OCn(Y in v1o/Do&Ro)Cn+1(Eo&Vo&Wo), and
(35) Cn+1(X∨Y)=∑o∈OCn((X∨Y) in v1
o/Do&Ro)Cn+1(Eo&Vo&Wo).
160160160160
Clearly, [(X∨Y) in v1
o] is equivalent to [(X in v1
o)∨(Y in v1
o)]. Thus,
(36) Cn+1(X∨Y)=∑o∈OCn((X in v1
o)∨(Y in v1o)/Do&Ro)Cn+1(Eo&Vo&Wo).
Because ~(X&Y) is a tautology, [~(X&Y) in v1
o] is also a tautology. Since [(X in v1
o)& (Y
in v1
o)] clearly entails [(X&Y) in v1
o], [(X in v1o)&(Y in v
1o)] is contradictory; hence, ~[(X
in v1
o)&(Y in v1
o)] is another tautology. By CADDn, Cn((X in v1
o)∨(Y in v1o)/Do&
Ro)=Cn(X in v1
o/Do& Ro)+Cn(Y in v1o/Do& Ro). Therefore,
(37) Cn+1(X∨Y)=
∑o∈OCn((X in v1
o)∨(Y in v1
o)/Do&Ro)Cn+1(Eo&Vo&Wo)=
∑o∈O[Cn(X in v1o/Do&Ro)+Cn(Y in v
1o/Do& Ro)]Cn+1(Eo&Vo&Wo)=
∑o∈OCn(X in v1
o/Do&Ro)Cn+1(Eo&Vo&Wo)+
∑o∈OCn(Y in v1o/Do&Ro)Cn+1(Eo&Vo&Wo)
=Cn+1(X)+Cn+1(Y).
In sum, GSJC+ preserves synchronic coherence from Cn to Cn+1. Note that I did not
depend upon the fact that this is a one-step updating. So GSR+ preserves synchronic
coherence from Cn to Cn+m for any m≥1.
Second, I argue that if B updates her credence functions in accordance with
GSJC+, then each of her credence functions diachronically coheres with the earlier
credence function. For this argument, let us consider any Cn, Cn-m∈∆. I will prove that Cn
161161161161
is related by GSJC+ to Cn-m under (i)-(iii);
71 once this proof is done, it will suffice to
argue that Cn diachronically coheres with Cn-m if Cn is related by GSJC+ to Cn-m.
Here goes the proof: Let Cn∈∆ be B’s credence function at tn. We want to show
that Cn is related by GSJC+ to Cn-m∈∆ for any m≥1. By (i), Cn is related by GSJC
+ to Cn-1,
Cn-1 is related by GSJC+ to Cn-2, …, and C1 is related by GSJC
+ to C0. By the earlier result,
Cn, Cn-1, ..., C0 are all synchronically coherent. By (1) and (2), GSJC+ and GSR
+ are
equivalent for those credence functions. Thus, Cn is related by GSR+ to Cn-1, Cn-1 is
related by GSR+ to Cn-2, …, and C1 is related by GSR
+ to C0. By the transitivity of GSR
+,
Cn is related by GSR+ to Cn-1, Cn-2, …, C0. By the equivalence, Cn is related by GSJC
+ to
Cn-1, Cn-2, …, C0. Done.
Next, consider this conditional claim: If Cn is related by GSJC+ to Cn-m, Cn
diachronically coheres with Cn-m, i.e., Cn is rationally related to Cn-m in the relevant sense.
How can I defend this claim? Remember my defense in the last chapter of GSJC as a
rational updating rule. According to it, if B sets her credences at tn by consulting her own
credal opinion at tn-m, then GSJC is the right way to do so. This is because, provided that
B has made a certain sequence of observations after tn-m but she still considers herself at
tn-m as an expert only lacking those observations, there exists a good argument that (a
principle entailing) GSJC captures the restriction that B’s self at tn-m imposes ? on B’s
credences at tn. If this is correct, this restriction will not only justify GSJC as a rule for
updating from Cn-m to Cn, but it also will justify it as a criterion of diachronic coherence
between Cn and Cn-m. Since the difference between GSJC and GSJC+ is ignorable here,
we have an argument for the wanted claim.
71 Since (i) is the antecedent of the wanted claim, it will be eventually discharged, but (ii) and (iii) will
remain to be substantial assumptions.
162162162162
Let me combine my discussions in the above two paragraphs: If an agent has
always updated her credences in accordance with GSJC+ (and (ii)&(iii) are satisfied),
then her present credence function Cn diachronically coheres with any past credence
function Cn-m of hers.
In sum, I have argued that GSJC+ satisfies the first requirement for the Final Rule.
This means that GSJC+ is not ruled out from being the general rule for de nunc updating,
at least not due to any failure to satisfy this first requirement.
G. Observational Exhaustiveness
In this section, I argue that first, an epistemic kernel version of GSJC+ prescribes an ideal
way to incorporate the totality of one’s observations into the present credal judgments,
and second, the original, transitional version of GSJC+ provides a good way to set one’s
credences in accordance with its epistemic kernel counterpart.
To begin, I formulate the epistemic kernel version of GSJC+: I assume that an
agent B had an initial credence function C0 at t0, a moment before she made any
observations. Hence, C0 can play the role of B’s kernel, as required for the formulation of
an epistemic kernel rule. Next, let Eo&Vo&Woo∈Ο be B’s general time-observation
partition from t0 to tn (constructed from Ω and Ψ), where Eo=&1≤k≤n(Eko at prevn-k),
Vo=&1≤k≤n(Vko at prevn-k), and Wo=(Wo at prevn). Given this partition, we can formulate
this rule for updating: For any tensed proposition X,
(GSJCE+
) Cn(X)=Σo∈ΟC0(X in vn
o/Do&Ro)Cn(Eo&Vo&Wo),
163163163163
where Do is the sequential de-indexicalization of Eo under Vo, and Ro is the sequential re-
indexicalization of Wo for the n epistemic moments earlier time (i.e., Do=&1≤k≤n(Eko in
vko) and Ro=Wo for any k∈1,...,n and any o∈Ο).
Why does GSJCE+
provide a good method for an agent to incorporate the totality
of her observations into her credal judgments? First, let’s think about a case in which the
agent has full knowledge about what observations she has made since t1 and what times it
has been since t0. In other words, she is fully certain at tn that she has observed E1, E
2, …,
En and that it has been w, v
1, v
2, …, v
n in those orders, for some <E
1, E
2, …, E
n>∈Ωn
and
<w, v1, v
2, …, v
n>∈Φn+1
. For such a case, GSJCE+
will provide this sub-principle: For any
tensed proposition X,
(GSSCE+
) Cn(X)=C0(X in vn/W&(E
1 in v
1)&(E
2 in v
2)&...&(E
n in v
n)).
Informally, this simply means that
(GSSCE+
) Cn(X)=
C0(X is true during vn/
it is w now&
E1 is true during v
1&
E2 is true during v
2&
...
&En is true during v
n).
164164164164
Observe two facts here: First, the above equation seems to capture a very
plausible way to incorporate the agent’s observations into her credal judgment at tn. For
what could be a more natural way to judge how probable X is at the present moment
(which the agent knows to be in vn) than to judge the probability of “X is true in v
n”
conditional on the conjoined initial truth of “it is w now,” “E1 is true during v
1,” “E
2 is
true during v2,” … and “E
n is true during v
n,” when the agent presently knows that this
conjunction has been confirmed by her observations so far? Second, and more
importantly, those conditions apparently capture all of the agent’s observations until tn.
Since the agent is assumed here to have observed nothing at t0, E1, E
2, …, and E
n exhaust
everything that she has observed until the present moment, tn.
Of course, one may be worried that the agent might not have a full knowledge
about what she has observed and/or what times it has been. For such cases, GSJCE+
provides a comparably intuitive strategy for epistemic kernel updating: If you do not
know what you have observed and/or what times it has been until now, first figure out
what credence you would assign to the target tensed proposition if you knew those facts,
and next take the weighted average of those credences with the weights being your
present credences in various scenarios about your observations and temporal locations
until now. Since each of the sequences of observations and times comprising those
scenarios exhaust all your observations since t1, I believe that GSJCE+
provides a
balanced way to incorporate the totality of your observations into the present credal
judgments.
165165165165
If so, it is easy to argue for the next main claim of this section. Remember the
earlier result that if an agent B has updated her credences in accordance with GSJC+, then
B’s present credence distribution Cn is related by GSJC+ to Cn-1, Cn-2, …, C0. Hence, Cn is
related by GSJC+ to C0 in the mentioned case. But it means that B’s credences at tn are set
in accordance with GSJCE+
.
This result suggests that if an agent updates her credences in accordance with
GSJC+
in the short run and repeats it, she comes to incorporate the totality of her
observations (since the initial moment) into her credences in the long run. In my opinion,
this is a big merit of GSJC+.
H. Filling the Gap
Remember that I used GSJC in the earlier chapters to solve the problem of Sleeping
Beauty. In this section, I first point out that we cannot do the same with GSJC+ because
of a new restriction on the general time-observation partition, and I then offer a solution
based upon a slightly modified version of GSJC+.
First, think about the following difference between GSJC and GSJC+: While
intervals of any size are allowed in the instances of GSJC, only intervals of a very small
size are allowed in those of GSJC+. For the members of Φ are such small intervals of
time that any better specification of one’s temporal location would be meaningless for
judging the relevance of the agent’s (de-indexicalized) observations to the (de-
indexicalized) target tensed proposition. One immediate consequence is that it is difficult
to apply GSJC+ to the cases in which time is specified in relatively coarse-grained units.
In particular, this means that we cannot solve the Sleeping Beauty problem by using
GSJC+. (Certainly, Monday and Tuesday do not belong to Φ.)
166166166166
To overcome this problem, I suggest yet another variant of GSJC, which I will
call “GSJC0.” The main difference between GSJC
+ and GSJC
0 is that time is specified by
using the members of Φ and Ψ in the former, but it is specified by using the unions or
disjunctions of their members in the latter.
To formulate GSJC0, let <∆,Ω,Φ,Ψ> be a model for an agent B’s credence
functions. Then, I will say that <∆,Ω,Θ,Ξ> is an extension of <∆,Ω,Φ,Ψ> iff it satisfies
the following conditions: (i) If an interval v belongs to Φ, then v also belongs to Θ; if
contiguous intervals v1,..., vn all belong to Φ, then ∪1≤j≤nvj also belongs to Φ; finally, no
other interval belongs to Θ. (ii) Ξ is a superset of Ψ such that if an interval w belongs to
Θ, then W or the tensed proposition that it is w now belongs to Ξ, and no other tensed
proposition belongs to Ξ. (Note that Φ can include intervals such as Monday and Tuesday;
correspondingly, Ξ can contain tensed propositions as those expressed by “it is Monday”
and “it is Tuesday.”)
Next, let Ep&Vp&Wpp∈P be an agent B’s general time-observation partition
from tn to tn+m, where Ep=&1≤k≤m(Ekp at prevm-k), Vp=&1≤k≤m(V
kp at prevm-k), and
Wp=&1≤k≤n+1 (Wkp at prevm+k-1) for each p∈P. But this time, we choose V
1p,...,V
mp and
W1
p,...,Wn+1
p not from Ψ but from Ξ. (In this case, I will say that Ep&Vp&Wpp∈P is
constructed from Ω and Ξ.) The rest of the formulation is similar: For any tensed
proposition X,
167167167167
(GSJC0) Cn+m(X)=Σp∈PCn(X in v
mp/Dp&Rp)Cn+m(Ep&Vp&Wp)
if Ep&Vp&Wpp∈P is optimal for X,
where Dp=&1≤k≤m(Ekp in v
kp) and Rp=&1≤k≤n+1 (W
kp at prevk-1). Note that we need the
explicit proviso of optimality because it is not guaranteed to be satisfied here.
Does GSJC0 meet the two requirements for the Final Rule? To answer this
question, we need to identify the logical relation between GSJC+ and GSJC
0. Let Cn and
Cn+m be B’s credence functions at tn and tn+m. Then, these claims are true:
(38) If Cn+m is related by GSJC0 to Cn, then Cn+m is related by GSJC
+ to
Cn, and
(39) If Cn and Cn+m are synchronically coherent and Cn+m is related by
GSJC0 to Cn, then Cn+m is also related by GSJC
+ to Cn.
(See APPENDIX D for the proofs.)
Once these facts are established, it is easy to argue for GSJC0’s satisfaction of the
two requirements for the Final Rule. First, suppose that the given agent, B, has always
updated her credences in accordance with GSJC0 until tn. Thus, Cn is related by GSJC
0 to
Cn-1, Cn-1 is related by GSJC0 to Cn-2, …, C1 is related by GSJC
0 to C0. By (38), Cn is
related by GSJC+ to Cn-1, Cn-1 is related by GSJC
+ to Cn-2, …, C1 is related by GSJC
+ to
C0. Since GSJC+ generates only synchronically coherent credence functions in this case,
so does GSJC0. Moreover, it was proven that, in this case, Cn is related by GSJC
+ to each
168168168168
of Cn-1, Cn-2, …, C0. By (39) and the already proved synchronic coherence of Cn, …, C1,
Cn is related by GSJC0 to each of Cn-1, Cn-2, …, C0. Therefore, GSJC
0 generates a
synchronically coherent credence function that diachronically coheres with the earlier
credence functions (if GSJC0 itself is used as the criterion of diachronic coherence).
Second, because Cn is related by GSJC0 to C0, Cn incorporates the totality of B’s
observations (if B observed nothing at t0).
Therefore, there exists a principle for updating that satisfies the two requirements
for the Final Rule. Furthermore, we can use that rule, GSJC0, in cases in which time is
only coarsely specified, as in the Sleeping Beauty problem.
I. Conclusion
In this chapter, I have argued that if there exists the Final Rule, or a rule that can always
be used by a rational agent to update her de nunc credences, it will satisfy the two
requirements discussed so far. Since GSJC+ (or, equivalently, GSJC
0) was shown to
satisfy those requirements, we have a promising candidate for the Final Rule.
169169169169
CHAPTER VI
CONCLUSION
A. Summary
So far, I have presented and defended GSJC, a new rule for de nunc updating. It has
various merits, the following being the most notable:
First, GSJC provides a convincing solution for the Sleeping Beauty problem: On
the one hand, waking up on Monday seems to be neutral between the coin’s landing
heads and landing tails. Hence, Cm(H/W& MON)=1/2. On the other hand, waking up on
Tuesday entails the coin’s landing tails. Thus, Cm(H/W&TUE)=0. If so, SB has to assign
the weighted average of ½ and 0 to the coin’s landing heads. These facts suggest that
Cm(H)∈(0,1/2). I find this line of reasoning to be convincing, and GSJC supports it
(under several plausible assumptions).
Second, GSJC applies to a wide range of cases, if not all: In the earlier chapters, I
first argued that SJC is the correct rule for de nunc updating, at least in some situations,
and then generalized the rule and argument together to show that GSJC is the generally
correct rule for de nunc updating. So GSJC applies to an agent’s credal transition from tn
to tn+m for any m≥1. Even if the agent is ignorant of what time it is now or what times it
had been until the time from which she is updating, GSJC applies to that credal transition
without a hitch.
Third, GSJC is coherent in many aspects: If you feed a synchronically coherent
prior credence function to that rule, it spits out another synchronically coherent credence
170170170170
function. Moreover, if you continue to update your credences in accordance with GSJC in
the short runs, then you come to have revised your credences in accordance with GSJC in
the long run. Plus, it allows you to incorporate all of your accumulated observations into
your present credal judgments.
In sum, GSJC provides an intuitive solution for the SB problem, it is applicable
to a wide range of cases, and it is coherent in many important aspects. These facts are
good reasons to accept GSJC as the general rule for de nunc updating.
Although these findings are nice achievements, three important issues are still
waiting for our discussion. The first concerns GSJC’s generality beyond de nunc
credences: As it is now, that rule is silent about how to update your de se credences in
general. The second concerns GSJC’s completeness as a rule for updating: If you update
in accordance with that rule, you need a predetermined credence distribution over your
general time-observation partition. Unfortunately, GSJC is silent about how to acquire
such a distribution. The third issue concerns GSJC’s complexity: Undoubtedly, it is a
highly complex rule. If there exists a simpler rule for de nunc/de se updating with the
same merits, isn’t it reasonable to prefer that simpler rule?
In this chapter, I will discuss these issues briefly, but I will refrain from full-
fledged discussions. My primary goals in writing this dissertation were, first, to present a
promising rule for de nunc updating and, second, to develop an argument to defend its
adequacy as the general rule for such an updating process. These goals are ambitious
enough for the first discussion of any updating rule, and so the more advanced
discussions will have to be saved for later papers or books.
171171171171
In the rest of this chapter, I will proceed in this order: In Section B, I will discuss
how to generalize GSJC for de se credences. In Section C, I will discuss in more detail
how to determine the credence distribution over the general time-observation partition. In
Section D, I will discuss the possibility of a simpler rule for de nunc or de se updating
with the same merits as GSJC’s, offering some reasons to be skeptical of that possibility.
B. Remaining Issue 1: Generalization for De Se Updating
At this point, it is natural to ask this question: “What is the correct rule for de se
credences in general?” In this section, I suggest a slightly modified version of GSJC as an
answer to this question.
Let’s begin. As before, let Eo&Vo&Woo∈O be the agent B’s time-observation
partition from tn to tn+m. This time, however, we allow each Eo&Vo&Wo to include any
centered-propositional letters, not just tensed-propositional ones. Then, for any centered
proposition X,
(GSJCde se) Cn+m(X)=Σ o∈O Cn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if
Eo&Vo&Woo∈O is optimal for X, and B is sure at tn+m that Cn was her own credence
function m epistemic moments ago,
where Do is the sequential de-indexicalization of Eo under Vo, and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time. To avoid confusion, I
will call the original, tensed version of GSJC “GSJCde nunc.”
172172172172
What are the differences between GSJCde nunc and GSJCde se? The first one is that
GSJCde se has centered-propositional letters in the places of tensed-propositional ones in
GSJCde nunc, and the second one is that GSJCde se has an additional proviso that the agent is
presently sure that Cn+m is her own credence function at the time from which she is
updating.
Set aside the second modification for a moment. Then, I have suggested simply
replacing tensed-propositional variables with centered-propositional ones. This
suggestion is attractive because scientists and philosophers tend to be conservative: They
always want to preserve as many elements of an established theory when they try to
generalize it for a broader range of cases. Nevertheless, we need to be careful because
this tendency sometimes misfires. For instance, when David Lewis suggested the de se
version of SC (hereafter: SCde se) as the correct rule for de se updating, he was obviously
being driven by the same kind of conservative tendency. Unfortunately, we know now
that SCde se is untenable.
However, I have some hunch that the expansion from GSJCde nunc to GSJCde se
will not misfire in a similar way. To understand why, note that we can conceive of three
types of beliefs: Those about “what this world is,” those about “what time it is now,” and
those about “who I am.” Without a defense, I assume that all beliefs are reducible to these
three types of beliefs or the combinations thereof.
The traditional theories of de dicto credences took only the first type of beliefs
into consideration. That made things easy. For think about this fact: You do not travel
from this world to that world, so the truth-values of the first type of beliefs do not change
through time. As a result, if you learn de dicto evidence E, then you can set your credence
173173173173
in a proposition X by feeding E directly into your previous conditional credence
function.72
In contrast, my theory of de nunc credence covers the second type of beliefs in
addition to the first. Unfortunately, this addition called for a drastic modification of the
traditional updating rules, which resulted in a much more complex rule, GSJCde nunc. The
reason was simple: As time flows, you travel from this time to that time. So the future
becomes the present, the present becomes the past, and the past becomes the farther past.
This fact demands complex techniques for shifting or translating your observations or
time-specifying propositions into the tensed propositions of the matching truth-values,
such as de-indexicalization and re-indexicalization. For example, if you learn “previously,
it was 9 AM,” you will set your present credence by using your previous credence
conditioned upon “it is 9 AM now.”
Now, we are talking about how to theorize the third type of beliefs, involving
“who I am,” in addition to the first and second types. Luckily, it is unlikely to be difficult
this time. Why? No one changes from this person to that person! So my hunch says that I
won’t need a complex technique for shifting in order to update the degrees of my beliefs
about “who I am.” For instance, if I newly learned (S) “I am the son of Sungki Kim,” I
will set my present credence by using my previous credence conditioned just upon S.
(Analogous to the degrees of your beliefs about who you are, are your friend’s beliefs,
your sister-in-law’s, etc.) If this hunch is correct, then it will be okay simply to replace
72 So Cn+1(X)=Cn(X/E). More generally, if your present experience directly affects your credences in Eis,
then your present credence in X is acquired by feeding Ei into your previous conditional credence function
and taking the weighted average. So Cn+1(X)=Σi∈ICn(X/Ei)Cn+1(Ei).
174174174174
tensed proposition letters with centered proposition ones within GSJC to get the general
rule for de se updating.
Still, we need to be careful. Think about this version of the SB problem: SB
problem 4. On Sunday, SB knows she will go through the following experiment: In the
next moment, a group of experimenters will put her to sleep. Then, they will toss a fair
coin. Case 1: The coin lands heads. In this case, she wakes up on Monday knowing that it
is Monday. Case 2: The coin lands tails. In this case, the experimenters duplicate SB
while she is sleeping. Then, they awaken her on Monday, and she knows that it is
Monday. What is her credence on Monday in the coin’s having landed heads?
In this version of the SB problem, when SB wakes up on Monday, she does not
know whether she is SB or the duplicate (hereafter: DUP). Let W be the centered
proposition expressed by “I am waking up today with the memory of SB’s until Sunday
as the last memory.” Given the analogy between the two versions of the problem, it is
likely that
(40) Cm(H)=
Cs(H/W is true of SB on Monday)Cm(W&I am SB)+
Cs(H/W is true of DUP on Monday)Cm(W&I am DUP)∈(0,1/2).
In this equation, SB’s evidence on Monday, W, is de-indexicalized to “W is true of SB on
Monday” or “W is true of DUP on Monday,” depending upon who she is. Clearly, this
fact contradicts my prior hunch that no shifting technique, such as de-indexicalization, is
175175175175
necessary involving the beliefs about “who I am” because no one changes from this
person to that person. What is happening here?
SB problem 4 shows that, although no one can change who she is, a rational
agent can sometimes be unsure about whether she is updating from her own previous
credence function or somebody else’s. From SB’s point of view on Monday, she is
updating from Cs; so, if she is SB, she is updating from her own previous credence
function, but if she is DUP, she is updating from somebody else’s previous credence
function, namely, SB’s.
Here is the general lesson: Usually, a shifting technique such as de-
indexicalization is unnecessary for de se updating because the agent will know the fact
that she is updating from her own prior credence function. However, there are rare cases
in which a rational agent is unsure of this fact. In such a case, some form of shifting
technique will be necessary involving the matter of “who I am.” The new proviso, “B is
sure at tn+m that Cn was her own credence function…” prevents GSJCde se from
erroneously applying to such cases.
In summary, it is acceptable in most of the targeted cases to generalize GSJCde
nunc for de se credences simply by replacing tensed-propositional letters with centered-
propositional ones, but a more complex method for updating will be necessary in some
rare cases. The proviso of GSJCde se is a safety device preventing its misapplication to
such cases.
176176176176
C. Remaining Issue 2: Credence Distribution over the Partition
At this point, a brilliant reader already will be asking this question: “How does an agent
determine her credence distribution over the general time-observation partition?” In this
section, I provide several possible answers.
To begin, I explain what the problem is. For comparison, think about the de dicto
version of JC (hereafter: JCde dicto). According to JCde dicto, an agent’s present credence in
proposition X ought to be the weighted average of her previous credences in X given Ei,
where the weights, the agent’s present credences in Eis, are somehow “directly affected
by” her present observations (Field 1978, p.361; Garber 1980, p.142).
This picture of “probability kinematics” indicates the existence of some relation
between the fineness of the observation partition and the agent’s perceptual power. For
example, suppose that you are watching a piece of cloth under a dim light and it appears
to be red or green to you, but you are unsure which color it is. So your observation
partition is R, G, where R is “this piece of cloth is red” and G is “this piece of cloth is
green.” Now you can imagine that if you had better eyesight, then you could distinguish
subtler colors, such as pinkish red, yellowish red, yellowish green, and bluish green. In
that case, your observation partition would be PR, YR, YG, BG (where PR is “this piece
of cloth is pinkish red,” etc.). Observe that this hypothetical observation partition is more
fine-grained than your actual observation partition. In general, the stronger your
perceptual power is, the finer-grained your observation partition is.
In this sense, your perceptual power sets the limit of the fineness of your
observation partition. This fact suggests that, if there exists a more fine-grained partition
than your current observation partition, then your credences in the members of that
177177177177
partition won’t be directly determined by your observations alone. For that partition will
be too fine-grained for its members’ credences to be determined only by your
observations.
If your perceptual power sets the limit of your de dicto observation partition’s
fineness in this sense, it is reasonable to think that the same is true of your de nunc
observation partition. Unfortunately, this means that at any moment, your credences in
your time-observation propositions cannot be directly determined by your observations.
For simplicity, let’s focus on SJC. According to that rule, Cn+1(X)=Cn(X in vj/Ei in vj)
Cn+1(Ei&Vj), where Cn and Cn+1 are your previous and present credence functions. To
calculate this value, you need to have a credence distribution over Ei&Vj<i,j>∈K at hand.
However, your perceptual power at tn+1 might be incapable of producing such a credence
distribution, if Ei&Vj<i,j>∈K is more fine-grained than Eii∈I.
So we have a problem: Although SJC requires that you have a credence
distribution over your time-observation partition at hand, the partition might be too fine-
grained for your perceptual experience alone to set the credence distribution over it. This
point, of course, generalizes to GSJC; for your general time-observation partition
normally will be more fine-grained than your observation partition.
How can we solve this problem? I do not have a fully developed solution yet, but
I have been considering three possible solutions. Let me outline them one by one. (Again
for simplicity, I will focus upon SJC only, but most of my points below will apply to
GSJC as well.)
First, your credence distribution over the time-observation partition may be
determined by appealing to a principle of indifference. I already discussed this approach
178178178178
in Chapter II, in which I criticized a principle of indifference endorsed by Adam Elga. In
a nutshell, his principle of indifference does not always support a particular credence
distribution over your time-observation partition (it does not if a member of the partition
spans over an uncountable number of worlds), nor is it guaranteed to be consistent (it
leads to a contradiction if the partition includes a possible world with an infinite but
countable number of subjectively indistinguishable centered worlds).
Nevertheless, it is still possible that (i) there exists a new version of the principle
of indifference, (ii) it provides a general recipe for setting your credence distribution over
your time-observation partition, and (iii) it is free from the problem discussed in the last
paragraph. If such a principle is found, it will produce a unique credence distribution over
your time-observation partition. Once such a credence distribution is provided, SJC can
do the rest of the job to calculate your credence function over the whole domain.
Sadly, various paradoxes tainted the reputation of the principle of indifference.
Here is the common structure of those paradoxes: The principle of indifference applies to
a partition, giving the same credence to its members, but in some situations, the same
possibilities are divisible into different partitions, leading to conflicting credence
assignments. Because of this type of problem, many philosophers have rejected the
principle of indifference. Recently, some philosophers have tried to revive it by
formulating a new principle of indifference invulnerable to the mentioned paradoxes.
(For instance, see Elga (2004), Mikkelson(2004), and White (ms.).) I personally am
skeptical of these attempts, but, to be fair, I say that it is too early to make the final
judgment. Of course, whether a principle of indifference can solve our problem depends
upon the success/failure of this general project.
179179179179
Second, if you cannot choose the right credence function over your time-
observation partition, perhaps it will be best not to choose only one of them. Instead, you
might rather have all possible credence distributions compatible with your observations
and previous credence function. Many philosophers suggest that, after all, it is unrealistic
to represent a human opinion with a single credence function because it is humanly
impossible to have such a precise opinion about anything. Instead, they claim that it is
better to represent a human opinion at a certain moment with a set of credence functions
(hereafter: representor) (Sturgeon, 2008; van Fraassen, 1990; Walley, 1991). Accordingly,
each proposition in the domain is assigned a set of real numbers (hereafter: vague
credence). Some advocates of vague credence think that an agent’s vague credence in a
proposition ought to include all compatible values of her credence in that proposition
with her data. I suspect that this idea provides a potential solution to our problem.
To see how this idea works, think about SB’s credal transition from s to m.
Remember that s is SB’s last conscious moment on Sunday, and m is the moment of her
waking up on Monday. Let CFSs be her representor at s. For any Cs∈CFSs, Cs(H)=1/2
because she knows on Sunday that the coin is fair. Let CFSm be her representor at m. At
m, her time-observation partition is W&MON,W&TUE. Since her observations cannot
uniquely determine the credences in W&MON and W&TUE, her vague credence at m in
W&MON ranges over (0,1).73
By SJC, Cm(H)=Cs(H on Monday/W on Monday)
73 Why not [0,1], (0,1] or [0,1)? I tend to think it is crazy that (*) CFSm includes a function which
completely rules out W&MON or W&TUE, since SB lacks any evidence logically contradicting either of
them. However, I admit that I do not have a ready answer to this question: “If (*) is crazy, is it not also
insane that CFSm includes a function that assigns 0.99999… to W&MON or W&TUE, given that she lacks
any evidence supporting either of them to a comparable degree?”
180180180180
Cm(W&MON) + Cs(H on Tuesday/ W on Tuesday)Cm(W&TUE)=1/2Cm(W&MON) for
any Cs∈CFSs and Cm∈CFSm. So her credence value set at m for H ranges over (0,1/2).
In my opinion, this solution is more attractive than the previous one (based on
the principle of indifference) for several reasons. First, as the name suggests, the theory
of vague credence is nothing more than one application of David Lewis’s general theory
of vagueness. If anybody is attracted to his theory of vagueness, she also will be attracted
to the notion of vague credence. Second, the idea of vague credence is motivated by
independent considerations. In reality, nobody can tell what single real number she
assigns to a proposition as the credence. Hence, if we adopt the notion of vague credence,
it will help to build a more realistic model of human credal opinions.
Third, if you are a die-hard subjectivist, you may wonder why there must be a
uniquely rational credence distribution, or even a set of credence distributions, over your
time-observation partition. To understand this point, think about these facts: In the
subjectivist tradition, a rational agent’s credal opinion is not assumed to supervene upon
her accumulated observational data; to establish supervenience, the agent’s initial
credence function needs to be included in the supervenience base. In other words, even if
perfectly rational agents A and B observe exactly the same data throughout their entire
lives, it is possible that A and B will have different credence functions at any time as long
as their initial credence functions were different.
How is it possible that A and B’s initial credence functions were different? Of
course, their initial credal judgments were made before any of their observations. So
there could not have been any a posteriori constraints on their initial credence functions.
Certainly, there must be some a priori constraints, but they will not be sufficient to make
181181181181
them have the same initial credence function. Therefore, this policy looks inevitable:
Allow A to have any initial credence function that she would like as long as it is
synchronically coherent and it satisfies other rational constraints. Mutatis mutandis for B.
Perhaps we can say the same thing about an agent B’s credence distribution over
her time-observation partition. As already emphasized, it is impossible to fully determine
this credence distribution if her time-observation partition is more fine-grained than her
observation partition. In such a case, perhaps this policy will be unavoidable: Allow B to
have any credence distribution over her time-observation partition that she would like as
long as it is synchronically coherent, it satisfies other a priori constraints, and it is
compatible with her credence distribution over her observation partition. Provided a
similar policy for the initial credence function, this laissez-faire policy is not so
implausible any longer.
So far, I have pointed out a problem regarding how to set your credence
distribution over your time-observation partition, necessary for using SJC, and I have
outlined three possible solutions to the problem. I do not pretend that these outlined
solutions are exhaustive; indeed, I do not find them to be fully satisfactory, and I
welcome any new solutions to this problem. If, however, we fail to find a better solution,
we can at least return to those outlined here as our fall-back positions. Of course, all these
problem and solutions are transferrable to GSJC and general time-observation partition.
D. Remaining Issue 3: The Possibility of a Rival Rule
Admittedly, GSJC is a complex rule. We all hate complex rules; they are hard to
understand and difficult to apply to real cases. We would of course prefer a simpler rule
for de nunc updating, all else being equal. However, in this section I argue that a rival
182182182182
rule satisfying the two conditions—“simpler” and “all else being equal”—will be hard to
find, if it exists at all.
The elements making GSJC a complex rule were introduced for some reasons,
after all. A GSJC-er will de-indexicalize her observations, which allows her to avoid the
problem of outdated conditional credence. (Review Chapters II and III.) A GSJC-er will
use the sequential de-indexicalization technique to deal with a sequence of observations,
which allows her to update from a previous, incorrect credence function. (Review
Chapter III.) Finally, a GSJC-er re-indexicalizes any de priori information she has, which
helps to overcome the problem of impoverished temporal knowledge. (Review Chapters
III and IV.) Clearly, these merits come with the cost of additional complexity. Still, the
merits exceed the cost.
Of course, if a simpler rule for de nunc updating enjoys all these merits, we will
favor such a rival rule over GSJC. I am, however, skeptical of this possibility. After all,
such a rival rule also will have to confront the problems mentioned in the last paragraph.
As I’ve shown, these problems can be solved if simplicity is abandoned, but the resulting
modification will not be much superior to GSJC in terms of simplicity. In sum, I am
concerned that a rival rule either will become equally complex after the necessary
modifications are made or will be unable to deal with some of the aforementioned
problems.
To illustrate this dilemma, I will discuss an alternative rule for de se updating as
a case study. Wolfgang Schwarz (ms.) suggests the following rule for de se updating: For
any centered proposition X, define >X to be that X will be true at the next epistemic
moment. For instance, if S is the centered proposition expressed by “I am watching the
183183183183
sun rise,” >S will be the centered proposition that S will be true at the next epistemic
moment. Using this notation, Schwarz formulates a very simple rule for de se updating:
Let X be any centered proposition, and let E be the totality of an agent B’s observations at
tn+1. Then,
(Shifted Conditioning) Cn+1(X)=Cn(>X/>E),
where Cn and Cn+1 are B’s credence functions at tn and tn+1. In words, an agent’s present
credence in a centered proposition X is equal to her previous credence in X’s truth at the
next epistemic moment, given E’s truth at the next epistemic moment, where E is the
totality of her present observations.
Clearly, Shifted Conditioning is a simpler rule than GSJC. However, can a
Shifted Conditionalizer deal with the aforementioned problems, which a GSJC-er can
handle easily? Not all of them.
First, let’s focus on the outdated conditional credence problem. Remember, this
problem: If you use Strict Conditionalization to set your present credence in the present
truth of X, then you come to set it by checking your previous conditional credence in the
previous truth of X.74
So, although you are trying to make a credal judgment of whether X
is presently true, you are doing so by using your previous credal judgment about whether
X was previously true, where the previous and present moments are different. If X is
irreducibly centered, this might be a problem because X’s truth-value might have changed.
74 For brevity, I omit the probabilistic antecedent of the previous conditional credence used for conditioning,
which is your present total evidence E.
184184184184
At first sight, Shifted Conditioning appears to be free from this problem. For, if
you update in accordance with Shifted Conditioning, you come to set your present
credence in the present truth of X by checking your previous conditional credence in the
truth at the then next moment of X.75
In the last sentence, “present” and “next” are co-
referential. As a result, you come to make a credal judgment about whether X is presently
true by using your previous credal judgment about whether X would be true at the then
next moment, just as it should be.
A problem occurs when the agent does not know that the present moment is next
to the moment from which she is updating. For example, when SB wakes up on Monday,
what will her credence that it is Monday be? According to Shifted Conditioning,
Cm(MON)=Cs(>MON/>W)=1.
In other words, her credence on Monday in its being Monday is equal to her credence on
Sunday night that the next epistemic moment would be on Monday, given that she would
wake up (with the memory of Sunday as the last memory) at the next moment. Since she
was sure on Sunday that it would be Monday in the next moment, the above instance of
Shifted Conditioning implies that waking up on Monday, she certainly knows that it is
Monday! This is crazy, because she is not in a position to know that it is Monday.
(One obvious escape route is to assume that actually there are two epistemic
moments next to the one on Sunday night—one on Monday and the other on Tuesday.
75 As before, I omit the probabilistic antecedent of the previous conditional credence used for Shifted
Conditioning, which is >E.
185185185185
For some curious reason, Schwarz rejects this potential solution.76
In any case, Shifted
Conditioning will have to be greatly modified in order to incorporate such a non-linear
structure of nextness among epistemic moments. And that modification will be done only
at the cost of increased complexity.)
What is the origin of this problem? On Sunday night, SB knew that the next
epistemic moment would be on Monday, whether the coin landed heads or tails. However,
when she actually wakes up on Monday, she cannot rule out that it is Tuesday. From her
point of view on Monday, if it is Tuesday, the present time is not next to the epistemic
moment on Sunday night. So she does not know whether or not the present moment is
next to the time from which she is updating. Still, Shifted Conditioning forces her to
update her credence in MON as if she knows that the present moment is next to the time
from which she is updating.
In a nutshell, updating in accordance with Shifted Conditioning can be a mistake
if the agent is not sure about how her present time is related to the time from which she is
updating. So Shifted Conditioning fails to solve the problem of outdated conditioning
credence adequately.
Second, Shifted Conditioning is not as versatile as GSJC because it is meant only
for the credal transition from the agent’s credence function at the previous epistemic
moment. Schwarz seems to be aware of this problem and provides a clue as to how to
remove this limitation: For any centered proposition X, he defines >nX to be the centered
proposition that X will be true at the n epistemic moments later time. For example,
suppose that going to bed, you fully expect to be awakened briefly during the night and to
76 Schwarz’s paper is unclear about this point. In fact, he clarified this point in a private email to me.
186186186186
be awakened by an alarm in the morning. Let WA be the centered proposition expressed
by “I am being awakened by the alarm.” Then, >2WA will be the centered proposition that
WA will be true at the two epistemic moments later time, which you presently know will
be the moment of waking up in the morning tomorrow. Using this expanded notation, I
formulate what I think to be a natural expansion of Shifted Conditioning: Let X be any
centered proposition, and let E1,…,E
m be such that for any k∈1,…,m, E
k is the totality
of her observations at tn+k. Then,
(Sequential Shifted Conditioning) Cn+m(X)=Cn(>mX/&1≤k≤m >kEk),
where Cn and Cn+m are B’s credence functions at tn and tn+m. According to this rule, if an
agent has observed E1, E
2,…, E
m, an agent’s present credence in X is equal to her
conditional credence at the m epistemic moments earlier time in [X’s truth at the m
epistemic moments earlier time], given [the conjunction of the truth at the k epistemic
moments later time of Ek for all k∈1,…,m]. I suppose this is the best way for Schwarz
to go. However, he does not explicitly endorse this extension of Shifted Conditioning,
and, even if he did, the cost would be increased complexity. (Also, remember that this
modification does not solve the outdated conditional credence problem adequately.)
Thus, as it is now, Shifted Conditioning cannot handle some of the problematic
cases with which GSJC has no problem. If you abandon its current simplicity, the thus-
modified rule may do better, but Schwarz will have to pay the cost of increased
complexity. Of course, I cannot completely rule out the possibility that somebody may
find a simpler rule for de nunc/de se updating, which somehow does not suffer from such
187187187187
problems. However, until I actually find such an alternative rule, I will remain skeptical
of that possibility.
E. Conclusion
In this dissertation, I have presented a series of rules for de nunc updating. I discussed
each of them critically, and, whenever I found a problem, I replaced an earlier rule with a
more general and plausible rule for updating. GSJC was the final product of this process.
I defended it by appealing to a variant of Gaifman’s expert principle, and I showed that it
has several highly desirable properties.
In addition, in this chapter I suggested that GSJC might be further generalized to
a rule for de se updating and complemented by some strategy for setting the credence
distribution over an agent’s general time-observation partition. Also, I explained why I
am skeptical about the possibility of any simpler rule for de nunc/de se updating enjoying
all the merits of GSJC. Given all these results and educated hunches, I believe that GSJC
is close to being the general rule for de se updating.
If this belief is correct, what will its general ramifications be? Traditionally, the
following elements have been considered to be the main elements of the theory of
subjective probability:
a) Non-negativity, Normality, and Additivity
b) Strict or Jeffrey-style conditionalization
c) Reflection Principle
d) Principal Principle
188188188188
These principles have been formulated in terms of de dicto credences. The important
question now is, “What happens if we take de se credences into consideration?”
Here is my conjecture: The entire theory of subjective probability needs
substantial modifications for the proper treatment of de se/de nunc credences, except the
synchronic axioms. Let me outline those modifications: In this dissertation, I have argued
that we need to replace SC/JC with GSJC. In his recent paper (2007), Adam Elga offered
a new variant of the Reflection Principle for agents who have lost track of what time it is.
I agree with his view that the original Reflection Principle needs to be modified to deal
with such cases, but I suspect that Elga’s suggestion for that modification is inadequate.
Plus, I believe that we need a new variant of the Principal Principle, which will connect
objective chance and de se credences. To my knowledge, no one previously has
mentioned the need to modify the Principal Principle in order to deal properly with de se
credences.
The theory of de se subjective probability is a vast territory consisting of
uncharted regions. In this dissertation, I explored one of its toughest parts, but the fun is
not over yet. There are still unexplored lands waiting for us.
189189189189
APPENDIX A
EQUIVALENCE BETWEEN GSJC- AND GSJC.
(=>) Suppose GSJC- and show GSJC. Let Eo&Vo&Woo∈O be an agent B’s general
time-observation partition from tn to tn+m, where Eo=&1≤k≤m(Eko at prevm-k),
Vo=&1≤k≤m(Vko at prevm-k), and Wo=&1≤k≤n+1(W
ko at prevm+k-1). By definition,
Eo&Vo&Woo∈O is B’s general time-observation partition from tn to tn+m over [t0,tn+m].
Let X be an arbitrary tensed proposition. By GSJC-,
(1) Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if Eo&Vo&Woo∈O
is optimal and sufficiently inclusive for X,
where Do is the sequential de-indexicalization of Eo under Vo and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time. For any o∈O,
Eo&Vo&Wo is its own abbreviation and W*o is the vacuous complement of Eo&Vo&Wo.
Clearly, Cn(X in vm
o/Do&Ro)=Cn(X in vm
o/Do&Ro&R*o), where Do is the sequential de-
indexicalization of Eo under Vo, Ro is the sequential re-indexicalization of Wo for the m
epistemic moments earlier time, and R*o is the vacuous sequential re-indexicalization of
190190190190
W*o for the m epistemic moments earlier time.77
Hence, Eo&Vo&Woo∈O is sufficiently
inclusive for X. Therefore,
(2) Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if Eo&Vo&Woo∈O
is optimal for X.
In other words, GSJC is true. Done.
(<=) Suppose GSJC and show GSJC-. Let Eo&Vo&Woo∈O be the general time-
observation partition from tn to tn+m. Given this partition, let Ep&Vp&Wpp∈P be B’s
general time-observation partition from tn to tn+m over [ti≤n,tn+m] i.e., for all p∈P, Ep=Eo,
Vp=Vo, and Wp=&1≤k≤n-i(Wko at prevm+k-1) for some i∈1,…,n+1. By GSJC,
(3) Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if Eo&Vo&Woo∈O
is optimal for X,
where Do is the sequential de-indexicalization of Eo under Vo and Ro is the sequential re-
indexicalization of Wo for the m epistemic moments earlier time.
77 Since Wp* is the complement of Eo&Vo&Wo for Eo&Vo&Wo, Wp*=&n+1≤k≤n+1(W
kp in prevm+k-1)=T, where
T is a tautology. By definition, Rp* is the re-indexicalization of Wp*, where Rp*=&n+1≤k≤n+1(Wkp in prevk-1).
Since no k satisfies n+1≤k≤n+1, Rp* is vacuously true, i.e., Rp*=T, where T is a tautology.
191191191191
Assume that Ep&Vp&Wpp∈P is optimal and sufficiently inclusive for X i.e.,
these conditions hold: First, for each p∈P, Cn(X in vm
p/Dp&Rp) is conditioned upon well-
specified temporal information, and the truth-value of X is invariant within vm
p and that of
Ekp is invariant within v
kp for any k∈1,…,m. Second, for each o∈O and p∈P, if
Ep&Vp&Wp is an abbreviation of Eo&Vo&Wo and Wp* is the complement, then Cn(X in
vm
p/ Dp&Rp)=Cn(X in vm
p/Dp&Rp&R*p), where Dp is the sequential de-indexicalization
of Ep under Vp, and Rp and Rp* are the sequential re-indexicalizations of Wp and Wp* for
the m epistemic moments earlier time. Given these assumptions, it suffices to show
(4) Cn+m(X)=Σp∈PCn(X in vm
p/Dp&Rp)Cn+m(Ep&Vp&Wp).
To show this, for each p∈P, let Op be the set of o∈O such that Ep&Vp&Wp is an
abbreviation of Eo&Vo&Wo. Clearly, Opp∈P is a partition of O. By this fact and (3),
(5) Cn+m(X)=Σp∈PΣo∈OpCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) if
Eo&Vo&Woo∈O is optimal for X.
Consider any p∈P. Then, Ep&Vp&Wp is an abbreviation of Eo&Vo&Wo for each o∈Op.
For each o∈Op, let Wo* be the complement of Ep&Vp&Wp for Eo&Vo&Wo. So for each
o∈Op, Eo&Vo&Wo=Ep&Vp&Wp&Wo*. On the one hand, Ep&Vp&Wp is incompatible
with Eo′&Vo′&Wo′ for any o′∉Op by its construction. Since Eo&Vo&Woo∈O is a partition,
this means that Ep&Vp&Wp entails Eo&Vo&Wo for some o∈Op. So Ep&Vp&Wp entails
192192192192
∨o∈Op Eo&Vo&Wo. On the other hand, Eo&Vo&Wo entails Ep&Vp&Wp as Eo&Vo&Wo=
Ep&Vp&Wp&Wo*. In sum, Ep&Vp&Wp is equivalent to ∨o∈Op Eo&Vo&Wo. Hence,
(6) Cn+m(Ep&Vp&Wp)=Σo∈OpCn+m(Eo&Vo&Wo).
Since Ep&Vp&Wp is an abbreviation of Eo&Vo&Wo and Wo* is the complement, Cn(X in
vm
p/Dp&Rp)=Cn(X in vm
p/Dp&Rp&R*o) by (iv), (where …). Since Ep= Eo, Dp= Do. Since
Wo= Wp&W*o, Ro= Rp&R*o. Hence,
(7) Cn(X in vm
p/ Dp&Rp)=Cn(X in vm
p/ Dp&Rp&R*o)=Cn(X in vm
p/Do&Ro).
Thus,
(8) if Eo&Vo&Woo∈O is optimal for X, Cn+m(X)=
Σp∈PΣo∈OpCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo)= (by (5))
Σp∈PCn(X in vm
p/Dp&Rp)Σo∈OpCn+m(Eo&Vo&Wo)= (by (7))
Σp∈PCn(X in vm
p/Dp&Rp)Cn+m(Ep&Vp&Wp). (by (6))
Remember that for all p∈P, Ep=Eo and Vp=Vo. By this fact and the given assumptions, for
each o∈O, Cn(X in vm
o/Do&Ro) is conditioned upon well-specified temporal information,
193193193193
and the truth-value of X is invariant within vm
o and that of Eko is invariant within v
ko for
any k∈1,…,m. In other words, Eo&Vo&Woo∈O is optimal for X. Done.
194194194194
APPENDIX B
EQUIVALENCE BETWEEN GSR+ AND GSJC
+
Let <∆,Ω,Φ,Ψ> be the model for B’s credences and Eo&Vo&Woo∈O be her general
time-observation partition from tn to tn+m constructed from Ω and Ψ, where Eo=
&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and Wo= &1≤k≤n+1(W
ko at prevm+k-1).
Let Cn\, Cn+m∈∆ be her credence functions at tn and tn+m. Then, these facts are provable:
(9) If Cn+m is synchronically coherent and related by GSR+ to Cn, then
Cn+m is also related by GSJC+ to Cn.
(10) If Cn is synchronically coherent and Cn+m is related by GSJC+ to
Cn, then Cn+m is also related by GSR+ to Cn.
To prove (9), suppose that Cn+m satisfies the standard axioms and that Cn+m(X/Eo&Vo
&Wo)=Cn(X in vm
o/Do&Ro) for each o∈O, where Do= &1≤k≤m(Eko in v
ko) and
Ro=&1≤k≤n+1(Wko at prevk-1). Then,
(11) Cn+m(X)=
Σο∈ΟCn+m(X&Eo&Vo&Wo)= (by Additivity)
195195195195
Σο∈ΟCn+m(X/Eo&Vo&Wo)Cn+m(Eo&Vo&Wo)= (def. of P(-/-))
Σο∈ΟCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo). (by supposition)
Done. Note that I used the assumed synchronic coherence of Cn+m in the second line.
To prove (10), suppose that Cn satisfies the standard axioms and that Cn+m(X)=
Σο∈ΟCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) for any X∈Ω. So Cn and Cn+m are functions
mapping Ω into R. (Note that we are not assuming yet that Cn+m is synchronically
coherent.) For each o∈O, construct a function Po mapping
Ωο=dfX&Eo&Vo&Wo|X∈Ω&o∈O into R as follows:
(12) Po(X&Eo&Vo&Wo)=dfCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo).
Let T be that 0=0, and F be that 0≠0. Thus, these facts are derivable:
(13) Po(Eo&Vo&Wo)=Po(T&Eo&Vo&Wo)=
Cn(T in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo)=Cn+m(Eo&Vo&Wo).
For (T in vm
o) is a tautology (e.g., “That 0=0 is true in June 2009” is a tautology).78
78 A careful reader may complain that if v
mo does not refer to an existing interval, (T in v
mo) might be false.
For example, “That 0=0 is true in June 2009” might be false if the world, including time itself, ceased to
exist at some time in 2008. If such a case is taken into consideration, then I suggest interpreting “X in ν” as
the tensed proposition that is true if and only if, if ν refers to an existing interval of time, then X is true
throughout ν, and, if ν does not refer to any existing interval of time, then X is a tautology.
196196196196
(14) Po(F)=Po(F&Eo&Vo&Wo)=
Cn(F in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo)=0.
For (F in vm
o) is a contradiction (e.g., “That 0≠0 is true in June 2009” is obviously a
contradiction). By (12) and (13),
(15) Po(X&Eo&Vo&Wo)/Po(Eo&Vo&Wo)=Cn(X in vm
o/Do&Ro).
(Also, remember that Po(Eo&Vo&Wo)=Cn+m(Eo&Vo&Wo)>0 by the definition of the
sequential time-observation partition.) Next, we define function P mapping Ω into R as
follows:
(16) P(X)=dfΣο∈ΟPo(X&Eo&Vo&Wo) for any X∈Ω.
Hence,
(17) P(X&Eo&Vo&Wo)=dfΣο∗∈ΟPo((X&Eo&Vo&Wo)&(Eο∗&Vο∗&Wο∗)).
(18) P(Eo&Vo&Wo)=dfΣο∗∈ΟPo((Eo&Vo&Wo)&(Eο∗&Vο∗&Wο∗)).
197197197197
Since Eo&Vo&Woo∈O is a partition, (X&Eo&Vo&Wo)&(Eο∗&Vο∗&Wο∗) is a
contradiction whenever o≠o∗, for any o, o*∈O. Thus,
(19) P(X&Eo&Vo&Wo)=Po(X&Eo&Vo&Wo). (by (17))
(20) P(Eo&Vo&Wo)=Po(Eo&Vo&Wo). (by (18))
So
(21) P(X&Eo&Vo&Wo)/P(Eo&Vo&Wo)=
Po(X&Eo&Vo&Wo)/Po(Eo&Vo&Wo)= (by (19) and (20))
Cn(X in vm
o/Do&Ro). (by (15))
By (12), (16), and supposition,
(22) P(X)=Σο∈ΟCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo)=Cn+m(X)
for any X∈Ω. Hence, P=Cn+m. By substitution in (21),
(23) Cn+m(X&Eo&Vo&Wo)/Cn+m(Eo&Vo&Wo)=Cn(X in vm
o/Do&Ro).
In other words, GSR+ is true. Done. Note that I depended upon the synchronic coherence
of Cn to prove (13) and (14).
198198198198
APPENDIX C
TRANSLATION BETWEEN TEMPORAL CONTEXTS
Remember that Eo&Vo&Woo∈O is B’s general time-observation partition from tn to tn+m,
Fp&Vp&Wpp∈P is B’s sequential time-observation partition from tn+m to tn+m+l, and
Gq&Vq&Wqq∈Q be B’s sequential time-observation partition from tn to tn+m+l, where
(24) Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k), and
Wo=&1≤k≤n+1 (Wko at prevm+k-1),
(25) Fp=&1≤k≤l(Fkp at prevl-k), Vp=&1≤k≤l (V
kp at prevl-k), and
Wp=&1≤k≤n+m+1(Wkp at prevl+k-1), and
(26) Gq=&1≤k≤m+l(Gkq at prevm+l-k), Vq=&1≤k≤m+l(V
kq at prevm+l-k), and
Wq=&1≤k≤n+1(Wkq at prevm+l+k-1).
I claim that
(27) Gq=&1≤k≤m(Eko at prevm+l-k)&&1≤k≤l(F
kp at prevl-k);
(28) Vq=&1≤k≤m(Vko at prevm+l-k)&&1≤k≤l(V
kp at prevl-k);
(29) Wq=&1≤k≤n+1(Wko at prevm+l+k-1) and Rq=&1≤k≤n+1(W
ko at prevk-1)=
Ro; and
199199199199
(30) Wp=&1≤k≤m(Vko at prevm+l-k)&&1≤k≤n+1(W
ko at prevm+l+k-1) and
Rp=&1≤k≤m(Vko at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1)=Vo&Wo,
tn+m+l, …,tn+m+l tn+m, …, tn+1 tn, …, t0
Time indexicals
used at tn+m+l
prev0,…,prevl-1 prevl, …, prevm+l-1 prevm+l, …, prevn+m+l
Time indexicals
used at tn+m
prev0, …, prevm-1 prevm, …, prevn+m
Time indexicals
used at tn
prev0, …, prevm
Figure 14: Referents of Indexicalsin Different Temporal Contexts. Each row shows B's
indexical specification of various times, from points of views at different times.
To see why, first, see Figure 14: In this figure, we find temporal moments
tn+m+l, ..., t0, to which we refer by using “tn+m+l,” ..., “t0.” However, the given agent, B,
may not have such non-indexical terms to refer to them with. Even in such a case, B can
still use indexicals terms to refer to those moments. In doing so, B can refer to the same
moment(s) by using different time indexicals at different times: For instance, she can
refer to the same moment tn by using “prevm+l” at tn+m+l, by using “prevm” at tn+m, and by
using “prev0” at tn. This means that she can ascribe the same tensed propositions to the
same moment, by believing different tensed propositions at different moments: For
example, let R be the tensed proposition that it is raining in White House now. Then, B
can ascribe R to tn, by believing [R at prevm+l] at tn+m+l, by believing [R at prevm] at tn+m,
and by believing [R at prev0] at tn.
200200200200
tn+m+l, …,tn+m+1 tn+m, …, tn+1 tn, …, t0
Time indexicals
used at tn+m+l
prev0,…,prevl-1 prevl, …, prevm+l-1 prevm+l, …, prevn+m+l
Observations
from tn+1 to tn+m+l
Gm+lq,…, Gm+1
q Gmq,…, G1
q
Time intervals
from t0 to tn+m+l
Vm+lq,…, Vm+1
q Vmq,…, V1
q W1q,…, Wn+1
q
Figure 15: Indexicals, Observations, and Intervals 1.This figure shows how B can ascribe
various observations and time intervals to the past epistemic moments.
Second, see Figure 15: Consider any Gq&Vq&Wq, (where Gq=&1≤k≤m+l(Gkq at
prevm+l-k), Vq=&1≤k≤m+l(Vkq at prevm+l-k), and Wq=&1≤k≤n+1(W
kq at prevm+l+k-1)).
Because of its construction, if B believes Gq&Vq&Wq at tn+m+l, she is ascribing
<Gm+l
q&Vm+l
q, …, G1q&V
1q,W
1q, …, W
n+1q> to <tn+m+l, ...,tn+1, tn, …, t0> at that moment.
79
Similarly, if she does not fully disbelieve Gq&Vq&Wq at tn+m+l, she is not completely
ruling out the ascription of <Gm+l
q&Vm+l
q, …, G1
q&V1
q,W1
q, …, Wn+1
q> to <tn+m+l, ...,tn+1,
tn, …, t0> at that moment. Since Cn+m+l(Gq&Vq&Wq)> 0, B is not fully ruling out that
ascription at tn+m+l.
Third, see Figure 16: Suppose, for reductio, that she ruled out this ascription of
<Gm
q&Vm
q, …, G1
q&V1
q, W1
q, …, Wn+1
q> to <tn+m, ...,tn+1, tn, …, t0> at tn+m. If so, she will
remember at tn+m+l that she has already ruled it out. In this case, B will rule it out at tn+m+l
79 Of course, I am not assuming that B knows at tn+m+l that the moments to which she is ascribing the tensed
propositions are <tn+m+l, …, tn+1, tn,..., t0>. Since she can use indexicals to refer to those epistemic moments,
the ability to identify those moments in non-indexical ways is unnecessary for such an ascription.
201201201201
as well, which contradicts the last paragraph. By reductio, she did not rule out the
ascription of <Gm
q&Vm
q, …, G1
q&V1q,W
1q, …, W
n+1q> to <tn+m, ...,tn+1, tn, …, t1> at tn+m+l.
tn+m+l, …,tn+m+1 tn+m, …, tn+1 tn, …, t0
Time indexicals
used at tn+m+l
prev0,…,prevl-1 prevl, …, prevm+l-1 prevm+l, …, prevn+m+l
Observations
from tn+1 to tn+m+l
Gm+lq,…, Gm+1
q Gmq,…, G1
q
Time intervals
from t0 to tn+m+l
Vm+lq,…, Vm+1
q Vmq,…, V1
q W1q,…, Wn+1
q
Time indexicals
used at tn+m
prev0, …, prevm-1 prevm, …, prevn+m
Observations
from tn+1 to tn+m
Emo,…, E1
o
Time intervals
from t0 to tn+m
Vmo,…, V1
o W1o,…, Wn+1
o
Figure 16: Indexicals, Observations, and Intervals 2. The dark area covers the indexicals,
observations, and times which have no counterpart in Eo&Vo&Wo.
Now, construct E∗&V∗
&W∗ so that E∗
=&1≤k≤m(Gkq at prevm-k), V
∗=&1≤k≤m(V
kq at prevm-k),
and W∗ =&1≤k≤n+1(W
kq at prevm+k-1). Due to its construction, if B fully disbelieved
E∗&V∗
&W∗ at tn+m, she would’ve ruled out the ascription of <G
mq&V
mq, …, G
1q&V
1q,
W1
q, …, Wn+1
q> to <tn+m, ...,tn+1, tn, …, t1> at that moment. Since she does not rule out that
ascription, B does not fully disbelieve E∗&V∗
&W∗ at tn+m. In other words,
Cn+m(E∗&V∗
&W∗)>0. However, remember that Eo&Vo&Woo∈O is B’s general time-
observation partition from tn to tn+m. By definition, the partition exhausts similarly
constructed tensed propositions whose credences at tn+m are strictly positive. Thus, there
must be o∈O such that E∗&V∗
&W∗=Eo&Vo&Wo. For this o∈O,
202202202202
(31) Gm
q=Em
o, …, G1
q=E1
o,
(32) Vm
q=Vm
o, …, V1q=V
1o, and
(33) W1
q=W1
o, …, Wn+1
q=Wn+1
o.
Fourth, see Figure 17: B does not rule out the ascription of <Gm+l
q&Vm+l
q, …,
Gm+1
q&Vm+1
q,Vm
q, …, V1q, W
1q, …, W
n+1q> to <tn+m+l, ..., t0> at tn+m+l. Now, construct
F∗&V∗&W∗ so that F∗=&1≤k≤l(Gm+k
q at prevl-k), V∗=&1≤k≤l (Vm+k
q at prevl-k), and
W∗=&1≤k≤m(Wkq at prevl+k-1)&&1≤k≤n+1(W
kq at prevm+l+k-1). Because of its construction,
if B fully disbelieves F∗&V∗&W∗ at tn+m+l, she would completely rule out the ascription
of <Gm+l
q&Vm+l
q, …, Gm+1
q&Vm+1
q, Vm
q, …, V1
q,W1
q, …, Wn+1
q> to <tn+m+l, ..., t0>
at that moment, which would contradict the above fact. Hence, Cn+m+l(F∗&V∗&W∗)>0.
However, Fp&Vp&Wpp∈P exhausts the similarly constructed tensed propositions whose
credences at tn+m+l are strictly positive.
tn+m+l, …,tn+m+1 tn+m, …, tn+1 tn, …, t0
Time indexicals
used at tn+m+l
prev0,…,prevl-1 prevl, …, prevm+l-1 prevm+l, …, prevn+m+l
Observations from
tn+1 to tn+m+l
Gm+lq,…, Gm+1
q Gmq,…, G1
q
Time intervals
from t0 to tn+m+l
Vm+lq,…, Vm+1
q Vmq,…, V1
q W1q,…, Wn+1
q
Observations from
tn+m+1 to tn+m+l
Fmp,…, F1
p
Time intervals
from t0 to tn+m+l
Vmp,…, V1
p W1p,…, Wm
p Wm+1p,…, Wn+m+1
p
Figure 17: Indexicals, Observations, and Intervals 3. The dark area covers the
observational data in Gq that have no counterparts in Fp.
203203203203
Hence, there exists p∈P such that F∗&V∗&W∗=Fp&Vp&Wp . For this p∈P,
(34) Gm+l
q=Flp, …, G
m+1q=F
1p,
(35) Vm+l
q=Vlp, …, V
m+1q=V
1p,
(36) Vm
q=W1
p, …, V1
q=Wm
p, and
(37) W1
q=Wm+1
p, …, Wn+1
q=Wn+m+1
p.
Therefore,
(38) Gq=&1≤k≤m+l(Gkq at prevm+l-k)= (by (26))
&1≤k≤m(Gkq at prevm+l-k)&&1≤k≤l(G
m+kq at prevl-k)= (by definition)
&1≤k≤m(Eko at prevm+l-k)&&1≤k≤l(F
kp at prevl-k), (by (31)&(34))
(39) Vq=&1≤k≤m+l(Vkq at prevm+l-k)= (by (26))
&1≤k≤m(Vkq at prevm+l-k)&&1≤k≤l(V
m+kq at prevl-k)= (by definition)
&1≤k≤m(Vko at prevm+l-k)&&1≤k≤l(V
kp at prevl-k), and (by (32)&(35))
(40) Wq=&1≤k≤n+1(Wkq at prevm+l+k-1)= (by (26))
&1≤k≤n+1(Wko at prevm+l+k-1). (by (33))
222204040404
Since Rq is the sequential re-indexicalization of Wq for the m+l epistemic moments
earlier time and Ro is the sequential re-indexicalization of Wo for the m epistemic
moments earlier time,
(41) Rq=&1≤k≤n+1(Wko at prevk-1)=Ro. (by definition)
Also,
(42) Wp=&1≤k≤n+m+1(Wkp at prevl+k-1)= (by (25))
&1≤k≤m(Wkp at prevl+k-1)&&1≤k≤n+1(W
m+kp at prevm+l+k-1)= (by definition)
&1≤k≤m(Vkq at prevm+l-k)&&1≤k≤n+1(W
kp at prevm+l+k-1)= (by (36)&(37))
&1≤k≤m(Vko at prevm+l-k)&&1≤k≤n+1(W
ko at prevm+l+k-1). (by (32)&(33))
Since Rp is the sequential re-indexicalization of Wp for the l epistemic moments earlier
time,
(43) Rp=&1≤k≤n+m+1(Wkp at prevk-1)= (by (25))
&1≤k≤m(Wkp at prevk-1)&&1≤k≤n+1(W
m+kp at prevm+k-1)= (by definition)
205205205205
&1≤k≤m(Vkq at prevm-k)&&1≤k≤n+1(W
kq at prevm+k-1)= (by (36)&(37))
&1≤k≤m(Vko at prevm-k)&&1≤k≤n+1(W
ko at prevm+k-1)= (by (32)&(33))
Vo&Wo. (by (24))
Done.
206206206206
APPENDIX D
EQUIVALENCE BETWEEN GSJC+ AND GSJC
0
Let Cn, Cn+m ∈∆ be B’s credence functions at tn and tn+m. Then, these facts are provable:
(44) If Cn+m is related by GSJC0 to Cn, then Cn+m is related by GSJC
+ to
Cn, and
(45) If Cn and Cn+m are synchronically coherent and Cn+m is related by
GSJC+ to Cn, then Cn+m is also related by GSJC
0 to Cn.
It is relatively easy to prove (44): Let <∆,Ω,Φ,Ψ> be a model for an agent B’s
credences and Eo&Vo&Woo∈O be her general time-observation partition from tn to tn+m
constructed from Ω and Ψ, where Eo=&1≤k≤m(Eko at prevm-k), Vo=&1≤k≤m(V
ko at prevm-k),
and Wo=&1≤k≤n+1(Wko at prevm+k-1). Suppose that Cn+m is related by GSJC
0 to Cn, and
show that Cn+m is related by GSJC+ to Cn. Hence, we want to show that
Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo) for any X∈Ω, where Do=
&1≤k≤m(Eko in v
ko) and Ro=&1≤k≤n+1(W
ko at prevk-1). Since Ψ⊆Ξ, Eo&Vo&Woo∈O was
constructed from Ω and Ξ. By supposition, Cn+m(X)=Σo∈OCn(X in vm
o/Do&Ro)
Cn+m(Eo&Vo&Wo) if (i) the truth-value of X is invariant within vm
o and that of Eko is
207207207207
invariant within vko for each k∈1,...,m and (ii) Cn(X in v
mo/Do&Ro) is conditioned upon
a well-specified temporal information, for each o∈O. Since X∈Ω and Eo&Vo&Woo∈O
was constructed from Ω and Ψ, (i) and (ii) are satisfied. Done.
It is more difficult to prove (45): Let <∆,Ω,Θ,Ξ> be an extension of <∆,Ω,Φ,Ψ>
and Ep&Vp&Wpp∈P be B’s general time-observation partition from tn to tn+m constructed
from Ω and Ξ, where Ep=&1≤k≤m(Ekp at prevm-k), Vp=&1≤k≤m(V
kp at prevm-k), and
Wp=&1≤k≤n+1(Wkp at prevm+k-1). Suppose that Cn and Cn+m are synchronically coherent
and Cn+m is related by GSJC+ to Cn, and show that Cn+m is also related by GSJC
0 to Cn.
To show this, let X be any member of Ω. It suffices to assume the satisfaction of (i) and
(ii) and show that Cn+m(X)=Σp∈PCn(X in vm
p/Dp&Rp) Cn+m( Ep&Vp&Wp), where Dp=
&1≤k≤m(Ekp in v
kp) and Rp=&1≤k≤n+1(W
kp at prevk-1). By construction, there exists
Ψ<k,p>⊆Ψ such that Vkp≡∨Ψ<k,p> for each k∈1,...,m and p∈P and there also exists
Ψ∗<k,p>⊆Ψ such that Wkp≡∨Ψ∗<k,p> for each k∈1,...,n+1 and p∈P. For each p∈P,
construct Eo&Vo&Woo∈O*p such that Eo=Ep, Vo=&1≤k≤m(Vko at prevm-k) for some
<V1
o,...,Vm
o>∈Ψ<1,p>×...×Ψ<m,p>, and Wo=&1≤k≤n+1(Wko at prevm+k-1) for some
<W1
o,...,Wn+1
o>∈Ψ∗<1,p>×...×Ψ∗<n+1,p>. Clearly, Eo&Vo&Woo∈O*p is a partition.
Consider arbitrary p∈P. On the one hand, ∨o∈O*pEo&Vo&Wo entails Ep&Vp&Wp.
To see this fact, it suffices to show that Eo&Vo&Wo entails Ep&Vp&Wp for any o∈O*p.
208208208208
So consider any o∈O*p. First, Eo clearly entails Ep. Second, Vo entails Vp. Why? For each
k∈1,...,m, (Vko at prevm-k) entails (V
ko at prevm-k), because V
ko∈Ψ<k,p> and so V
ko entails
∨Ψ<k,p>, which is the same as Vkp. It clearly follows that &1≤k≤m(V
ko at prevm-k) entails
&1≤k≤m(Vkp at prevm-k). Third, Wo entails Wo. Why? For each k∈1,...,n+1, (W
ko at
prevm+k-1) entails (Wko at prevm+k-1), because W
ko∈Ψ∗<k,p> and so W
ko entails ∨Ψ∗<k,p>,
which is the same as Wkp. It clearly follows that &1≤k≤n+1(W
ko at prevm+k-1) entails
&1≤k≤n+1(Wkp at prevm+k-1).
On the other hand, Ep&Vp&Wp entails ∨o∈O*pEo&Vo&Wo: By the construction of
Eo&Vo&Woo∈Op, there exists some o∈O*p such that Eo=Ep, Vo=&1≤k≤m(Vko at prevm-k)
for some <V1
o,...,Vm
o>∈Ψ<1,p>×...×Ψ<m,p>, and Wo=&1≤k≤n+1(Wko at prevm+k-1) for some
<W1
o,...,Wn+1
o>∈Ψ∗<1,p>×...×Ψ∗<n+1,p>. So Vko entails ∨Ψ<k,p> for each k∈1,...,m.
Similarly, Wko entails ∨Ψ∗<k,p> for each k∈1,...,n+1. Since V
kp≡∨Ψ<k,p> and
Wkp≡∨Ψ∗<k,p>, V
ko entails V
kp for each k∈1,...,m and W
ko entails W
kp for each
k∈1,...,n+1. Clearly, (Vko at prevm-k) entails (V
kp at prevm-k) for each k∈1,...,m and
(Wko at prevm+k-1) entails (W
kp at prevm+k-1) for each k∈1,...,n+1. Thus, Vp&Wp entails
Vo&Wo. Since Ep clearly entails Eo, Ep&Vp&Wp entails that Eo&Vo&Wo for some o∈O*p.
Hence, Ep&Vp&Wp entails ∨o∈O*pEo&Vo&Wo.
209209209209
Since p was arbitrarily chosen from P, Ep&Vp&Wp is equivalent to
∨o∈O*p(Eo&Vo&Wo) for each p∈P. For each p∈P, construct Op⊆O*p such that
Cn+m(Eo&Vo&Wo)>0 for any o∈Op. Define O to be ∪p∈POp. Hence,
(46) Cn+m(Eo&Vo&Wo)>0 for any o∈O,
(47) Σo∈OCn+m(Eo&Vo&Wo)=
Σp∈PΣo∈OpCn+m(Eo&Vo&Wo)= (by the construction of O)
Σp∈PCn+m(∨o∈Οp(Eo&Vo&Wo))= (by Additivity)
Σp∈PCn+m(∨o∈O*p(Eo&Vo&Wo))= (by the construction of each Op)
Σp∈PCn+m(Ep&Vp&Wp)=1, and (by the above equivalence)
(48) Σo∈OCn(Do&Ro)>0. (by (46))
To understand how (48) derives from (46), suppose, for reductio, that Σo∈OCn(Do&Ro)=0.
Then, Cn(Do&Ro)=0 for any o∈O. It means that at tn, B completely rules out the
possibility that [E1
o will be true in v1
o,E2o will be true in v
2o, …, E
mo will be true in v
1o]
and [it is w1o now, it was w
1o at the one epistemic moment earlier time, …, it was w
n+1o at
the n+1 epistemic moment earlier time]. At tn+m, B remembers that she ruled out this
possibility at the m epistemic moments earlier time. Acknowledging that m epistemic
210210210210
moments have passed, she will rule out the possibility that [E1o&V
1o was true, E
2o&V
2o
was true, …, Em
o&Vm
o is true now] and [it was w1o at the m epistemic earlier time, it was
w1o at the m+1 epistemic moment earlier time, …, it was w
n+1o at the m+n+1 epistemic
moment earlier time]. So Cn+m(Eo&Vo&Wo)= 0. However, this contradicts (46). By
reductio, Σo∈OCn(Do&Ro)>0. From (46)-(48), it follows that Eo&Vo&Woo∈O is B’s
sequential time-observation partition from tn to tn+m over [t0,tn+m] constructed from Ω and
Ψ. Since we supposed that Cn+m is related by GSJC+ to Cn,
(49) Cn+m(X)=Σο∈ΟCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo).
Focus upon the value of Cn+m(Eo&Vo&Wo). For any o∈O such that o∈Op,
(50) Cn+m(Eo&Vo&Wo)=
Σp∈PΣo∈OpCn+m(Eo&Vo&Wo)= (by the construction of O)
Σp∈PΣo∈O*pCn+m(Eo&Vo&Wo)= (by the construction of Op)
Σp∈PCn+m(∨o∈O*p(Eo&Vo&Wo))= (by Additivity)
Σp∈PCn+m(Ep&Vp&Wp). (Ep&Vp&Wp≡ ∨o∈O*p(Eo&Vo&Wo))
Focus upon the value of Cn(X in vm
p/Dp&Rp). For any p∈P and o∈Op,
211211211211
(51) Cn(X in vm
p/Dp&Rp)=
Cn(X in vm
p/&1≤k≤m(Ekp in v
kp)&&1≤k≤n+1(W
kp at prevk-1))= (by definition)
Cn(X in vm
o/&1≤k≤m(Ekp in v
kp)&&1≤k≤n+1(W
ko at prevk-1)).
For Cn(X in vm
p/Dp&Rp) was assumed to be conditioned upon well-specified temporal
description and wko⊆w
kp, for each k∈1,...,n+1. For any p∈P and o∈Op,
(52) Cn(X in vm
o/&1≤k≤m(Ekp in v
kp)&&1≤k≤n+1(W
ko at prevk-1))=
Cn(X in vm
p/&1≤k≤m(Eko in v
kp)&&1≤k≤n+1(W
kp at prevk-1))=
Cn(X in vm
o/&1≤k≤m(Eko in v
ko)&&1≤k≤n+1(W
kp at prevk-1)),
because the truth-value of X is invariant within vm
p and that of Eko (=E
kp) is invariant
within vkp, and v
mo⊆v
mp and v
ko⊆v
kp, for any k∈1,...,m. Hence,
(53) Cn+m(X)=
Σο∈ΟCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo)= (by (49))
Σο∈Οp&p∈PCn(X in vm
o/Do&Ro)Cn+m(Eo&Vo&Wo)= (by construction of O)
Σο∈Οp&p∈PCn(X in vm
p/Dp&Rp)Cn+m(Ep&Vp&Wp)= (by (50)-(52))
212212212212
Σp∈PCn(X in vm
p/Dp&Rp)Cn+m(Ep&Vp&Wp). (simplification)
Therefore, Cn+m is related by GSJC+ to Cn. Done.
213213213213
BIBLIOGRAPHY
Alchourron, C.E., Gardenfors, P., Makinson, D. (1985). On the Logic of Theory Change.
Journal of Symbolic Logic, 50, 510-530.
Bradley, D. (2003). Sleeping Beauty: a note on Dorr's argument for 1/3. Analysis, 63 ,
266-268.
Bricker, P. (ms.) Realism without Parochialism. An unpublished manuscipt.
Connee, E., & Feldman, R. (2004). Evidentialism. Oxford: Oxford University Press.
Dietrich, F., & List, C. (forthcoming). The Aggregation of Propositional Attitudes:
Towards a General Theory. In T. S. Gendler, Oxford Studies in Epistemology.
Oxford: Oxford University Press.
Dorr, C. (2002). Sleeping Beauty: in defence of Elga. Analysis, 62 , 292-296.
_______ (2000). Self-locating Beliefs and Sleeping Beauty Problem. Analyssis, 60 , 143-
147.
_______ (2004). Defeating Dr. Evil with Self-locating Belief. Philosophy and
Phenomenological Research, 69 , 383-396.
Elga, A. (2007). Reflection and Disagreement. Nous, 41 , 478-502.
Field, H. (1978). a Note on Jeffrey Conditionalization. Philosophy of Science, 45 #3 ,
361-367.
Gaifman, H. (1988). A Theory of Higher Order Probabilities. In B. Skyrms, & W. Harper,
Causation, Chance, and Credence (pp. 191-220). Dortrecht: Kluwwer Academic
Publishers.
Garber, D. (1980). Field and Jeffrey Conditionalization. Philosophy of Science, 47 #1 ,
142-145.
Hajek, A. (2003). What Conditional Probability Could Not Be. Synthese, 137 #3 , 273-
323.
Hall, N. (2004). Two Mistakes about Credence and Chance. Australian Journal of
Philosophy, 82 , 93-111.
214214214214
Jeffrey, R. (1984). Baysianism with a Human Face. In J. Earman, Testing Scientific
Theories: Minnesota Studies in the Philosophy of Science (pp. 133-156).
Minneapolis: University of Minnesota Press.
_______ (1990). The Logic of Decision. London: University of Chicago Press.
Katsuno, H., & Mendelzone, A. O. (1992). On the Difference between Updating a
Knowledge Base and Revising it. In P. Gardenfors, Belief Revision (pp. 183-203).
Cambridge: Cambridge University Press.
Kierland, B., & Monton, B. (2005). Minimizing Inaccuracy for Self-Locating Beliefs.
Philosophy and Phenomenological Research, 70, 384-395.
Lewis, D. K. (1979). Attitudes De Dicto and De Se. The Philosophical Review, 88 , 513-
543.
_______ (1986). A Subjectivist's Guide to Objective Chance. In R. Jeffrey, Studies in
Inductive Logic and Probability, Vol. II. (pp. 263-293). Berkeley: University of
Chicago Press.
_______ (2001). Sleeping Beauty: Reply to Elga. Analysis, 61 , 171-176.
Meacham, C. J. (2008). Sleeping Beauty and Dynamics of De Se Beliefs. Philosophical
Studies, 138 , 245-269.
_______ (forthcoming). Unraveling Tangled Web: Continuity, Internalism, Uniqueness
& Self-Locating Belief. In T. S. Gendler, & J. Hawthorne, Oxford Studies in
Epistemology, Vol. 3. Oxford: Oxford University Press.
Mikkelson, M. J. (2004). Dissolving the Wine/Water Paradox. British Journal of
Philosophy of Science, 55, 137-145.
Pollock, J., & Cruz, J. (1986). Contemporary Theories of Knowledge. Lanham: Rowman
& Littlefield.
Prior, A. N. (2003). Papers on Time and Tense. Oxford: Oxford University Press.
Rescher, R., & Urquhart, A. (1971). Temporal Logic. New York: Springer-Verlag.
Schwarz, W. (ms.). Changing Minds in a Changing World. An unpublished manuscript.
Sturgeon, S. (2008). Reason and the Grain of Belief. Nous, 42 #1 , 139-165.
Talbott, W. J. (1991). Two Principles of Bayesian Epistemology. Philosophical Studies,
62 , 135-150.
215215215215
van Fraassen, B. (1984). Belief and the Will. Journal of Philosophy, 81 , 235-256.
_______ (1990). Figures in a Probability Landscape. In A. Gupta, & M. Dunn, Truth or
Consequences (pp. 345-356). Dortrecht: Kluwer.
Walley, P. (1991). Statistical Reasoning with Imprecise Probabilities. Chapman & Hall.
Weatherson, B. (2005). Should We Respond to Evil with Indifference? Philosophy and
Phenomenological Research, 70 , 614-635.
Weintraub, R. (2004). Sleeping Beauty: a simple solution. Analysis, 64 , 8-10.
White, R. (ms.) Evidential Symmetry and Mushy Credence. An unpublished manuscript.